What this PR does / why we need it:

Adds SavedDatasetLanceStorage, so an existing Lance dataset can be recorded as
a saved dataset without moving or rewriting it.

This is the path for a dataset produced outside Feast. create_saved_dataset
takes a RetrievalJob, executes it and persists the result, which is wrong for
an embedding dataset that a Ray or Spark job already wrote to Lance.
BaseRegistry.apply_saved_dataset records metadata only, and that is what this
storage class is for:

storage = SavedDatasetLanceStorage(LanceFormat(version=7), uri="s3://bucket/embeddings")
registry.apply_saved_dataset(
    SavedDataset(
        name="poi_embeddings_v7",
        features=[],
        join_keys=[],
        storage=storage,
        tags={"model": "clip-vit-l-14", "recall_at_10": "0.9137"},
    ),
    project,
)

The format must pin a version

SavedDatasetLanceStorage rejects an unpinned LanceFormat. A saved dataset
addressing whatever version is latest at read time does not identify the rows it
was created from, so a measurement recorded against it stops being attributable.
version is preferable to tag, since a tag can be moved afterwards.

Note that LanceSource defaults table_format to an unpinned LanceFormat
rather than None, so from_data_source on a plain source raises rather than
silently recording a drifting reference.

Dispatch

Lance is the fourth class to claim the custom_storage oneof field, after
postgres, clickhouse and couchbase. to_proto therefore goes through
SavedDatasetStorage._custom_storage_proto (#6988) so the payload records the
writing class; otherwise from_proto would resolve by import order.

Verified that a Lance payload reads back as SavedDatasetLanceStorage with all
four claimants registered, across several import orders:

[postgres clickhouse couchbase lance]  claimants=4  -> SavedDatasetLanceStorage  version=7
[lance postgres clickhouse couchbase]  claimants=4  -> SavedDatasetLanceStorage  version=7
[clickhouse lance couchbase postgres]  claimants=4  -> SavedDatasetLanceStorage  version=7

LanceSource is also registered in _DATA_SOURCE_TO_SAVED_DATASET_STORAGE, so
SavedDatasetStorage.from_data_source resolves it.

Which issue(s) this PR fixes:

N/A. Follows #6988, which made the shared custom_storage field safe for a
fourth claimant.

Checks

• I've made sure the tests are passing.
• My commits are signed off (git commit -s)
• My PR title follows conventional commits format

Testing Strategy

• Unit tests
• Integration tests
• Manual tests
• Testing is not required for this change

sdk/python/tests/unit/infra/data_sources/test_saved_dataset_lance_storage.py,
14 tests: round trip by uri and by namespace+table, the payload recording the
storage class, an unpinned format refused, a wrong-typed format refused, uri and
table exclusivity, a table without namespace_impl refused, to_data_source
and from_data_source preserving the pin, a non-Lance source refused, the
_DATA_SOURCE_TO_SAVED_DATASET_STORAGE registration, and dispatch with all four
custom_storage claimants present.

Against master: ruff check and ruff format --check clean, and mypy feast
reports the same 24 pre-existing errors as origin/master, none added. The
adjacent merged suites (tests/unit/infra/data_sources/,
tests/unit/test_table_format.py, tests/unit/test_saved_dataset_storage_dispatch.py)
show an identical set of failures to clean master, all in
test_mlflow_dataset_source.py from a missing local dependency.