docs: dataset and experiment features, API reference fixes, and two new guides - #1118

…ew guides

## Summary

Brings the dataset and experiment docs up to date with the next platform release, fixes dataset API reference pages that did not work as written, adds an Experiments API reference, and adds two guides that existed only in the internal docs. Every statement was re-checked against the hosted product source; where hosted and self-hosted behave differently, the wording is true for both.

## Merge timing

Merge this with the next platform release, not before. Most of these pages describe dashboard controls and API fields that reach hosted users in that release:

- Dataset splits: the Use splits guide, the split filter and Manage splits, Export as training set, and the dataset splits API
- The More > Download… export chooser, the typed blank grid, the media and persona cell editors, and the dashboard upload formats (CSV, JSON, JSONL)
- Hugging Face shown as unavailable in the dashboard
- The experiment wizard steps, split selection, Re-run with changes, failed-cell reruns, default evaluations, and the experiment limits and statuses
- The named, typed column schema for dataset creation in the API, and column roles
- `split_ids` and `append_only` in the Experiments API, and the split-scoped SDK experiment example

The other fixes (the cell-update method and clearing behavior, the accepted column types, the upload size limit, and the two new guides) are also true today, but they ship in this one change.

## Dataset API reference

- **Update cell value:** the method is POST, not PUT. The response envelope is corrected, as is the error for a missing cell (404). `new_value` is documented: omitting it, or sending null or an empty string, clears the cell, and JSON and array columns take serialized JSON strings.
- **Add static column:** the example now uses an accepted column type. The accepted types are listed, and the optional column role is documented.
- **Create dataset:** documents the optional named, typed column schema, with its defaults and limits.
- **Upload dataset:** the size limit is 25 MiB. Excel is accepted by the REST API but not by the dashboard upload.
- **New: Dataset splits:** documents split create, update and delete, membership changes, and the bounded JSON export.

## Experiments API reference (new)

Adds an overview plus pages for creating, reading, updating, re-running, re-running cells, stopping, deleting and comparing experiments. These pages cover the contract that both hosted and self-hosted share.

## Dataset guides

- **New: Use splits:** live, named subsets of rows, with filtering and Export as training set.
- **Manage datasets:** the More > Download… export chooser (CSV, JSON, JSONL, Parquet and the OpenAI formats), and the editors for media and persona cells.
- **Create a dataset and Add rows:** the typed blank grid, and the dashboard upload formats (CSV, JSON and JSONL).
- **Overview:** Hugging Face import is marked as unavailable in the dashboard.
- **Limits and data types:** adds experiment limits and statuses.

## Experiments

- **Run an experiment:**
  - split selection in the Data step;
  - Re-run with changes, which adds a variant and keeps earlier results;
  - re-running failed cells;
  - default evaluations.
- **Compare prompts cookbook:** the steps now follow the current Configuration → Data → Evaluations → Review & run wizard.

## SDK

- **Datasets:** cell updates pass `new_value` explicitly, with None or null to clear. Uploads use the platform client's multipart call. Python reconnects by exact name. Paged reads are shown for exports.
- **New: Experiments:** platform client create, get and list, including split-scoped creation.

## New guides

- **Share a trace:** restricted and public links. It keeps the warning that a public link exposes the trace's contents.
- **Fetch prompts in production:** Python's process cache and its stale fallback, and how TypeScript returns templates.

## Checks

- Five independent read-only reviews found the gaps. Each page group was written separately and then given a correctness review against the product source.
- A final pass found no broken internal links, balanced MDX on every changed page, and navigation entries for every new page.