feat(table): add RowDelta API for atomic row-level mutations

[laskoviymishka]

Adds Transaction.NewRowDelta() — Go equivalent of Java's BaseRowDelta. Commits data files and delete files (position or equality) in one atomic snapshot. This is needed for row-level mutations: an UPDATE becomes an equality delete for the old row + append of the new row, both in one commit.

Resolves #602.

API

rd := tx.NewRowDelta(snapshotProps)
rd.AddRows(dataFile1, dataFile2)
rd.AddDeletes(posDeleteFile, eqDeleteFile)
rd.Commit(ctx)

Operation type picked automatically: data-only → append, deletes-only → delete, both → overwrite.

Validation

• Delete files require format version >= 2
• Equality deletes must have non-empty EqualityFieldIDs referencing existing schema columns
• Content types checked: no data files in AddDeletes, no delete files in AddRows

Known limitations

• No conflict detection for concurrent writers — documented in the type comment
• Uses fast-append producer (no manifest merging)

What's tested

The interesting ones:

• Commit data + position deletes, check snapshot summary has added-data-files=1, added-delete-files=1, operation is overwrite
• Commit equality deletes, check added-equality-delete-files shows up in summary
• Read back manifests after commit, verify there's one data manifest and one delete manifest with correct content types in entries
• Two RowDeltas on same transaction (batch1 append, batch2 append+delete), verify cumulative total-data-files
• v1 table rejects delete files with clear error
• Equality delete file without field IDs → error
• Equality delete file with field ID 999 (not in schema) → error

The round-trip integration test:

1. Write 5 rows as real Parquet, append to table
2. Write a position delete file targeting positions 1 and 3, commit via RowDelta
3. Scan the table back — get 3 rows, verify IDs are [1, 3, 5] (beta and delta gone)

This covers the full path: write parquet → RowDelta commit → scan with position delete filtering applied.

What's left to do

This PR covers the commit API. Remaining work for full DML support:

• Equality delete file writing — a writer that produces Parquet files with PK-only schema and EntryContentEqDeletes content type. The RowDelta API already accepts them, but there's no convenient writer yet.
• Equality delete reading — the scanner currently errors with "iceberg-go does not yet support equality deletes" (scanner.go:415). Needs: collect eq delete entries during scan planning, match to data files by partition + sequence number, apply hash-based anti-join during Arrow reads.
• Conflict validation — validateFromSnapshot, validateNoConflictingDataFiles, etc. Java's Flink connector skips most of this for streaming, so it's not blocking for CDC use cases.