Persist configured distance_metric in dataset metadata (not just runtime kwarg)

[dcfocus]

Motivation

#74 / #77 add a configurable vector-search distance_metric (l2 / cosine /
dot) as a runtime option on ContextStoreOptions. Unlike embedding_dim
— which is recovered automatically from the persisted Arrow schema on open
(embedding_dim_from_schema, crates/lance-context-core/src/store.rs) — the
distance metric is not persisted. A caller must pass the same
distance_metric on every open, and if they forget, the store silently falls
back to the default l2 and ranks results differently from how the dataset was
intended to be queried.

Proposal

Persist the chosen metric in the dataset so it round-trips without being
re-specified.

• On create, write the metric into Lance schema metadata (the same
  mechanism already used for blob encoding, e.g. the
  lance-encoding:blob keys in store.rs), e.g.
  lance-context:distance_metric = "cosine".
• On open, read it back and use it as the store's metric. An explicitly passed
  distance_metric option may override (or, preferably, error on mismatch,
  mirroring the existing embedding_dim mismatch check in
  open_with_options).
• Default remains l2 for datasets created before this change (absent key).

Scope

• Core write path (schema metadata on create) + read-back on open.
• Reuse/mirror the embedding_dim mismatch-validation pattern.
• Tests: create with cosine, reopen without passing the option, confirm
  ranking still uses cosine; legacy dataset (no key) still defaults to l2.

Non-goals

• Changing the metric set or the ranking math (handled in Make vector-search distance metric configurable (cosine/dot-product; currently hardcoded L2) #74).
• ANN/vector indexing.

Relationship

Follow-up to #74 / #77. embedding_dim already persists via the column schema,
so this issue is specifically about the metric setting.