perf: lazy + projected payload reads (#116)

[dcfocus]

Summary

Closes #116. Reads previously fully materialized binary_payload (and text_payload / embedding) for every row on every list / search / get — so a metadata listing or a vector search over multi-modal records pulled every image's bytes into memory. This adds column projection on the read path plus on-demand blob fetch.

Independent of #115 — it operates on the existing inline payload columns, so it can land on its own.

What changed

Core (lance-context-core)

• ReadProjection { text, binary, embedding } — default loads everything (backward compatible), with metadata_only() / without_binary() helpers.
• batch_to_records now tolerates those columns being projected out (omitted fields decode as None).
• ContextStore::list_filtered_projected and search_filtered_projected read only the requested columns; the existing *_with_options methods delegate with the full projection (no behavior change). Search always reads the embedding internally to score, then drops it from results if projection.embedding is false.
• ContextStore::get_blob(id) -> Option<Vec<u8>> fetches one record's bytes via a projected, id-filtered scan.

Python

• Context.list(..., include_binary=True, include_embedding=True) and Context.search(..., include_binary=..., include_embedding=...).
• Context.get_blob(id) -> bytes | None.
• New pyo3 args default to true, so existing callers are unaffected.

lean = ctx.list(filters={"content_type": "image/png"}, include_binary=False)  # no bytes
img  = ctx.get_blob(lean[0]["id"])                                            # fetch on demand
hits = ctx.search(query_embedding, include_binary=False)                      # ranked, no bytes

Tests

• core: projection drops binary/embedding but keeps metadata; get_blob fetches on demand (and returns None for missing); search projection still ranks correctly.
• python: test_lazy_payload.py — list/search exclude binary, get_blob round-trip.

Checks

• cargo test -p lance-context-core --lib — 98 passed
• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings — clean
• ruff format --check, ruff check, pyright — clean

Acceptance criteria (#116)

• list / search can avoid materializing binary_payload (and embedding)
• Projection / include_binary flag controls inclusion (default preserves current behavior)
• get_blob(id) fetches a single record's bytes on demand
• Tests cover projection + on-demand fetch
• REST include_binary params + blob GET route — deferred follow-up
• Lance take_blob/BlobFile zero-copy lazy reads for blob-encoded columns — deferred follow-up

Notes

• I defaulted projection to load-all (non-breaking) rather than flipping the default to exclude binary, since silently dropping bytes from existing list/get results would be a breaking change. Flipping the default can be revisited as a separate, deliberate change.

Closes #116