arena: add public allocation interface

[edsiper]

Summary

• expose raw request-lifetime allocation through cfl_arena_malloc() and cfl_arena_calloc()
• add cfl_arena_memdup() and cfl_arena_strndup() helpers
• add an options-based constructor with optional allocator callbacks and bounded geometric chunk growth
• retain fixed-size chunks as the default and preserve the existing constructors
• document alignment, zero-size, overflow, failure, reset, and lifetime semantics
• extend unit, installed-consumer, and benchmark coverage

The arena remains opaque, and the API additions are source- and ABI-additive. Existing CFL arena users retain their current fixed-size allocation policy unless they explicitly select a larger maximum chunk size. This PR only provides the general interface; it does not migrate downstream consumers.

Validation

• cmake -S . -B build-public-arena -DCFL_TESTS=On -DCFL_BENCHMARKS=On
• cmake --build build-public-arena -j8
• ctest --test-dir build-public-arena --output-on-failure (34/34 passed)
• ASan/UBSan test suite with leak detection (34/34 passed)
• focused arena tests under Valgrind (no leaks or errors)
• installed-consumer configure, build, link, and execution
• downstream builds/tests: cmetrics (20/20), ctraces (4/4), cprofiles (4/4)
• Fluent Bit default build, including processor plugins
• git diff --check

Benchmark support

The arena benchmark now compares fixed-size chunks with an optional bounded-growth mode. This is intended to support representative consumer-side measurements, including OTLP protobuf rendering, without claiming workload-specific performance from the synthetic benchmark alone.

[edsiper]

Superseded by #87 using the project topic branch naming convention.