Firestore DML fixes

[MarkDuckworth]

Follow-up to #9118. This aligns the DML and literals pipeline APIs with the existing Pipeline API conventions, fixes several encoding and validation bugs in the literals stage, and documents how insert/upsert choose their target documents and what atomic does.

Fixes API problems that the proposal review missed, before the API ships. It also fixes literals inputs that the SDK previously dropped silently, crashed on, or encoded in a way the backend rejected.

Changes

API shape

• literals(documents) + literals(options), with documents required in LiteralsStageOptions. Why: other stages take their required argument directly and also offer an options overload that carries it. The backend rejects a literals stage with no documents, so documents shouldn't be optional.
• update: removed the Record<string, Expression> | Map overload and added update(options: UpdateStageOptions) with an optional transformedFields. Why: no other stage accepts a map of fields; they take AliasedExpressions. Every stage should also have an options overload. This removed overload was not in the API proposal or tests.
• delete() + delete(options: DeleteStageOptions). Why: every stage should have an options overload, e.g. for rawOptions. This follows the database() precedent for stages with no required parameters.
• insert() + insert(options), replacing a single optional-options signature. Why: this matches the database() precedent and the other DML stages.
• upsert(), upsert(field, ...fields) and upsert(options); removed upsert(fields[], options?). Why: this matches the variadic convention used by select(), addFields(), etc. The removed overload could pass fields in two places, and the code had to merge them.
• Removed the deprecated transforms option from UpsertStageOptions. Why: it was a pre-release name for additionalFields, added and renamed within feat(firestore): add DML stages, literals source, and atomic execution option to Node SDK pipelines #9118. It was never released in Node or Web, so there's nothing to deprecate.
• literals is only on PipelineSource now, not on Pipeline. Why: literals is a source stage. It was never tested in a non-first position, and the Web SDK only exposes it on PipelineSource.
• The Delete and Update stages now forward rawOptions. Why: they were silently dropping them, unlike every other stage.
  literals encoding and validation
• FieldValue sentinels (serverTimestamp(), increment(), arrayUnion(), delete(), ...) in literal documents now throw Function literals() called with invalid data. FieldValue.serverTimestamp() can only be used with update() and set() (found in field <path>). Why: a pipeline has no way to represent sentinels. Before this change they were silently dropped, nested data was lost, and a document containing only a sentinel failed with an opaque 13 INTERNAL serialization error. The message matches the Web SDK. Validation is now recursive, which also covers nested user data. currentTimestamp() is the server-side alternative to serverTimestamp().
• Nested objects and arrays in literal documents are automatically wrapped in map() / array() expressions. Why: the backend only evaluates expressions in the top-level fields of a literal document, so {a: {b: add(...)}} and {arr: [add(...)]} failed with Value type is not supported: FUNCTION_VALUE. Each document is still sent as a top-level mapValue. Its fields are converted with valueToDefaultExpr, the same conversion used for expression arguments and rawStage().
• undefined in literal documents is rejected unless ignoreUndefinedProperties is enabled. Why: this keeps the previous serializer behaviour, which the new encoding would otherwise have bypassed by silently dropping undefined.

[gemini-code-assist]

Code Review

This pull request refactors several Firestore pipeline stages—including literals, delete, update, insert, and upsert—to accept options objects, updating their signatures, type definitions, and tests accordingly. The review feedback focuses on improving robustness and decoupling: it suggests moving the selectablesToMap conversion into the UpdateStage constructor to keep the Pipeline class clean, adding safe default fallbacks to prevent runtime crashes when options are undefined, and strengthening validation in the literals stage to ensure each document is a plain object.