Add allocation_binary_trees benchmark

[danvinci]

Adds allocation_binary_trees - a memory-allocation / traversal benchmark - implemented similarly across all 14 languages.

Each iteration builds a complete binary tree of depth 14 (16,383 nodes), traverses it to compute a node-count checksum, and frees it. 25 iterations per timed window, NITER=5, taking minimum.
Each language uses its idiomatic allocation strategy: malloc/free in C, new/GC in Java/Go/JS/Swift/Scala, structs in Julia/Rust, dicts/lists in Python/Numba/R/Octave, tables in Lua, allocatable types in Fortran.

The existing 8 benchmarks are mostly numeric / BLAS-leaning; this one probes a dimension where allocation strategy and GC dominate timing.

• One commit per language, plus README + benchmark-order/colors meta updates
• No new external dependencies
• Single-threaded execution conventions match PR Standardize benchmark methodology and enforce single-threaded execution #112
• In-language assertion: bt_sum == 25 * (2^14 - 1) * NITER verifies each tree builder

Test plan

• Verified locally on macOS arm64: all 14 implementations compile and emit lang,allocation_binary_trees,time_ms
• All in-language assertions pass
• CI workflow passes for all 14 language benchmarks
• report job (aggregation + deploy) passes
• Documentation build succeeds (Documenter preview deployed)

---

Drafted with assistance from Claude Opus 4.7.