fix: return actionable errors for binary/string offset overflow

[Copilot]

Reading or compacting string / binary data could still hit Arrow offset-overflow panics when a single batch exceeded 2GiB, leaving users with an opaque failure and no clear mitigation. This change converts those cases into normal Lance errors with guidance to reduce batch size or move to large_string / large_binary.

• Decode overflow handling

  • Add explicit size checks before narrowing offsets in legacy binary/string decode paths.
  • Return NotSupported errors instead of allowing Arrow internals to panic on oversized batches.
  • Preserve existing Arrow errors for non-overflow failures (for example, invalid UTF-8 or malformed offsets).

• User-facing diagnostics

  • Replace the generic “offset overflow” / “too many lists” style failure with actionable messaging for binary-like columns.
  • Point users to the two real mitigations:
    • lower batch_size / LANCE_DEFAULT_BATCH_SIZE
    • convert the column to large_string / large_binary
  • For list decoding, clarify which prefix of rows fits and which requested range overflows.

• Compaction-adjacent behavior

  • Ensure the same decode-time protection applies when compaction reads string/binary data through these paths, so the failure is surfaced as a normal Lance error instead of a panic.

• Representative error shape

  Could not create array with more than 2GiB of string/binary data in a single batch.
  Please reduce the batch_size, set LANCE_DEFAULT_BATCH_SIZE to a smaller value,
  or convert the column to large_string/large_binary.
  

• Fixes Offset overflow errors can be confusing for users #2775

[github-actions]

ACTION NEEDED
Lance follows the Conventional Commits specification for release automation.

The PR title and description are used as the merge commit message. Please update your PR title and description to match the specification.

For details on the error please inspect the "PR Title Check" action.

[lance-gatekeeper]

✅ Gate recommendation: approve.

Both decode paths now reject oversized normalized spans before narrowing offsets, while fallible Arrow construction preserves unrelated failures. This keeps normal decoding and the storage format unchanged, and avoids the fixed-length UTF-8 cost reported by the always-64-bit alternative.