feat(realtime): support two-stage scan and read

[lxy-9602]

Purpose

Linked issue: #158

This PR extends the real-time append read path introduced in #163 to support separate scan and
reader creation stages.

The main changes are:

• capture immutable memory views during scan planning and register them in RealtimeContext;
• add version, snapshot id, partition-bucket, committed offset, memory upper offset, and an opaque
  read-view ticket to RealtimeSplit;
• allow ReadContext to share the same RealtimeContext used by the writer and scanner;
• resolve the ticket when creating a reader and validate that its partition-bucket and offset range
  match the split;
• use RealtimeReader to keep the memory view alive while disk and memory data are being read;
• release abandoned or completed read views through a background release queue, avoiding large
  segment destruction on the query thread;
• preserve an already-created reader across commit, refresh, and memory-segment reclamation.

A real-time split ticket is currently single-use. Once reader creation consumes the ticket, retrying
the same reader task requires creating a new scan plan and RealtimeSplit. Tickets are also
process-local and must be resolved by the RealtimeContext that created them.

Tests

Added or updated coverage for:

• pinning, resolving, and releasing a read-view ticket;
• expiration and cleanup of abandoned read-view tickets;
• rejection when the same RealtimeSplit ticket is reused;
• reading from a reader created before commit and refresh after its memory segment is reclaimed;
• reading a pinned plan after writer close;
• updated Arrow memory-indexer reader ownership.

API and Format

This PR changes public APIs:

• adds ReadContextBuilder::WithRealtimeContext;
• adds read-view ticket lifecycle APIs to RealtimeContext;
• adds RealtimeReader, which keeps a MemReadView alive for the reader lifetime;
• changes MemIndexer::CreateQueryReaders to return RealtimeReader instances.

There is no change to the Paimon table storage format, data-file format, snapshot format, or commit
protocol.

RealtimeSplit carries versioned real-time metadata, but cross-process split serialization is not
implemented in this PR. Its opaque ticket is currently resolved only inside the originating process.

Documentation

Generative AI tooling

Generated-by: OpenAI Codex (GPT-5)

[zjw1111]

Concurrent consumption of the same real-time split needs to be prevented.

[SteNicholas]

This class is compiled under the repository-wide -fvisibility=hidden, but realtime_context_test.cpp and realtime_write_inte_test.cpp call its out-of-line methods while their executables link against paimon_shared. Those methods are therefore private/hidden in the shared library (for example Cast, GetOrCreateMemIndexer, and ResolveReadView), so the test link will fail with undefined symbols. Please either export the test-used class/symbols, link this implementation into those tests, or exercise it through public APIs.

[SteNicholas]

This consumes the ticket before the public TableRead::CreateReader(vector<Split>) call has succeeded as a whole. With two real-time splits, split 1 can release its ticket here and split 2 can then fail (for example, a plugin error or ticket expiry); the vector overload returns an error, but retrying that failed call now fails on split 1 because its ticket is gone. Please make the ticket claims/consumption rollbackable across the vector overload, or defer consumption until all per-split readers have been created, and add a two-split regression test.