Fall back when the encoder buffer pool is exhausted

[shin4141]

What this fixes

When all pooled EdgeBuffer instances are in use, borrow() gets null from pool.poll() and the next encode throws at reset(). Allocate a new EdgeBuffer only for that exhausted-pool case. This is the follow-up invited by the maintainer on #505.

Why it matters

Concurrent encodes beyond poolSize can fail even though an independent buffer can encode the value.

Review scope

Two codec-java files: a fallback in AbstractEdgeEncoder.borrow() and one concurrency regression test. The pooled fast path and release behavior stay the same.

Evidence

• On main at afa7143f1d8a0d16623eb5de3e963706ff819bba, the regression reproduces the second encode's NullPointerException at useAsByteArray:408.
• With this patch, overlapping encodes use distinct buffers and both outputs match unpooled controls.
• :codec-java:test: 68 passed; :codec-java:spotlessCheck: passed.
• :core:test --tests '*V2BulkLoadAggregationTest*': 1 passed.
• Full engine suite was not run for this follow-up and remains unverified.

AI Assistance

• This change was written largely with AI assistance.
  • Tool / model: Codex (GPT-6).

[em3s]

Thanks @shin4141 — clean fix, and the latch test shows the exhausted-pool case directly. LGTM, merging.