`src/megatron/bridge/data/hf_processors/` exports three pure-function
dataset processors (squad, gsm8k, openmathinstruct2) used by the
`default_*_config` helpers in finetune_utils.py. Each one is a
dict-in / dict-out function with no I/O and no tokenizer dependency.

They had **functional tests** under
`tests/functional_tests/test_groups/data/hf_processors/` but **zero
unit tests** — meaning every regression in the input/output formatting
required a GPU container CI slot to catch.

This PR adds unit-test coverage so processor-format regressions get
caught at L0-unit-test time. 22 tests across 4 classes:

`TestProcessSquadExample` (5):
  - basic example produces documented `Context: ... Question: ...
    Answer:` formatting; output is the FIRST answer in the answers
    list; original_answers preserves all alternatives
  - single-answer example produces a valid output
  - input strictly ends with bare `Answer:`
  - tokenizer arg is a no-op (None and a sentinel produce identical
    output) — locks in the documented contract
  - missing required field surfaces a clear KeyError

`TestExtractFinalAnswer` (5):
  - extracts value after `####`, strips whitespace, returns full
    answer stripped when no delimiter, returns empty string when
    `####` is followed by whitespace, uses LAST split when `####`
    appears multiple times

`TestProcessGsm8kExample` (4):
  - basic example: `Question: ... Answer:` formatting; output is the
    full chain-of-thought; original_answers contains ONLY the
    extracted final numerical answer
  - no-`####` answer flows through `_extract_final_answer` correctly
  - input strictly ends with `Answer:`
  - tokenizer arg is a no-op

`TestProcessOpenMathInstruct2Example` (5):
  - basic example: `Problem: ... Solution:` formatting; output is the
    generated solution; original_answers wraps expected_answer in a
    one-element list
  - input strictly ends with `Solution:`
  - expected_answer is preserved verbatim (no stripping)
  - tokenizer arg is a no-op
  - missing required field surfaces KeyError

`TestProcessorOutputContract` (3, parametrized):
  - cross-processor invariant: every processor returns a dict with
    `input` (str), `output` (str), and `original_answers` (list[str]
    with at least one element)

Tests-only — no production code changes. Locks in the formatting
contract every recipe that uses these processors implicitly depends on.

Signed-off-by: lonexreb <[email protected]>