chore(release): consolidate fixes and prepare Instructor 1.17.0 - #2597

`pdf_to_anthropic` encoded any PDF with unset `data` by calling
`requests.get(str(pdf.source))`. That branch is only reachable for
non-HTTP sources, because remote `http(s)` URLs return earlier as a URL
document block. Every source that could reach it (a local path, `gs://`,
`s3://`) is therefore one `requests` cannot resolve, so it raised
`MissingSchema`/`InvalidSchema` rather than encoding the document.

`PDF.source` is documented as "URL, file path, or base64 data of the
PDF", and constructing `PDF(source=Path(...))` directly leaves `data`
unset. Building a PDF from a local path and sending it to Anthropic thus
always failed, even though the file was readable on disk.

Read local files from disk instead, and raise the same clear `ValueError`
that `pdf_to_openai` raises when no usable source is present. This also
removes the last raw `requests` call left in a provider multimodal
module after remote media fetching was centralized in #2569.

The existing coverage test asserted the broken behaviour by mocking
`requests.get`, and additionally failed on Windows because it compared a
`Path` against a hardcoded POSIX string. It now writes a real file to
`tmp_path`, and two regression tests cover the local-path and
missing-source cases.

Remove the runtime dependency on legacy HTTPX during automatic client
construction. Permit OpenAI 3 while preserving OpenAI 2 support on
Python 3.9, and cover both synchronous and asynchronous paths.

Refs: #2553

@async_field_validator and @async_model_validator only attached metadata
via setattr; nothing ever read it, so declared async validators were
silently never invoked and validation errors never triggered a reask.

After a non-streaming parse succeeds, retry_async_v2 now recursively runs
declared async validators (including on nested models) and raises the
existing AsyncValidationError into the reask path on failure. A sync
client given a response_model that declares async validators now raises
ConfigurationError immediately instead of silently skipping validation,
since a sync client can never await them.

Fixes #2528.

Fixes the 12 formatting-lint failures the scheduled Weekly Tests job reports
on main (black/ruff per-block via pytest_examples):

- validation.md: drop unused Field import, prefer double quotes
- distillation.md: trailing comma in RAW-format dict literal
- logging.md: remove extra blank line black collapses
- reask_validation.md: typing_extensions.Annotated -> typing.Annotated,
  extra blank line
- batch.md: drop unused BatchProcessor/extract_results imports, quote style
- hooks.md: drop unused *args, underline unused lambda args, black line
  wrapping

All 141 tests in tests/docs/test_concepts{,_advanced}.py pass locally.

The Gemini multimodal post showed its sample output as `Recomendations(`,
but the model class is defined and passed as `Recommendations`, so the
snippet did not match the code directly above it. The Gemini integration
page linked to that post as "Using Geminin To Extract Travel Video
Recomendations".

Also corrects spelling in prose and example prompts elsewhere: ambigious,
convinient, useable, recieves, targetted, succint, contrain (constrain),
Lanuage, Declaritive, occuring, probabilites, genearted, indentifying,
numberical, Janurary, Februray, concatinate, procide, duplcates, runnign,
serveral, roomate, isntance, neeed and possibily, plus two doubled words
("a a world class", "of the the signed date").

`tests/docs/test_posts.py` lints every code block under `docs/blog/posts` with
black via pytest-examples. 34 of them fail on `main`, which is why `weekly-tests`
and the `release` job's test step are both red.

This fixes 16 of the 34. Two were genuinely broken Python -- a reader copying
them gets a SyntaxError:

  bad-schemas-could-break-llms.md   " lass AnswerWithIntermediateCalculations"
                                    -> "class ..."  (missing the c)

  jinja-proposal.md                 "topic": "...7890"     <- no comma
                                    "banned_words": [...]
                                    two adjacent string literals, so the dict
                                    key after them is a syntax error

Once those parse, black can format them, so the rest of the change is what
pytest-examples' own formatter produces: joined lines that now fit, and blank
lines black adds or removes between definitions.

In jinja-proposal.md the formatter also restored `#> ` on an output comment that
had drifted to `# >`. `#> ` (with the trailing space) is pytest-examples'
`comment_prefix`; `# >` is not matched as an output marker.

Deliberately not touched: introducing-structured-outputs.md. Its block contains
a bare `#> ` marker standing for an empty printed line, and that form cannot
survive formatting -- black strips the trailing space, the result no longer
matches `comment_prefix`, and black then rewrites it to `# >`, which both breaks
the marker and leaves it inconsistent with the `#> ...` lines under it. Writing
`# >` into the docs to satisfy the linter seemed worse than leaving it, so it is
in the group below.

The remaining 18 need a call I did not want to make unilaterally, since they are
published posts:

  - empty ```python blocks where install instructions appear to have been lost
    (multimodal-gemini.md, generating-pdf-citations.md) -- the prose introduces
    a snippet that is not there
  - pseudo-code fenced as python, e.g.
    `client.create(..., response_model=type[T]) -> T` in version-1.md, which is
    a type signature rather than a statement
  - fragments that were never standalone modules, e.g. a bare `@field_validator`
    in chain-of-density.md

For those the options are to complete the snippet, re-fence as ```text (only
fences starting with `py`/`{.py` are collected), or exclude older posts from the
lint. Happy to follow up with whichever you prefer.