perf: reduce compile and IO overhead

[alvgeppetto-debug]

Reduce compile and IO overhead in the training loop.

Changes:

• ACCUM_STEPS configurable via ANE_ACCUM_STEPS env var (default 10). Higher values mean fewer exec() restarts and better effective throughput.
• MAX_COMPILES configurable via ANE_MAX_COMPILES env var (default 100). Allows tuning for different hardware/OS versions.
• IOSurface pooling: reuse freed surfaces by size instead of creating new ones. Avoids repeated IOSurfaceCreate/CFRelease on every recompile cycle. Pool capacity: 128 surfaces with swap-remove lookup.
• Applied to both tiny_train.m and train_large.m codepaths.

Both make train and make train_large compile cleanly on macOS.

[alvgeppetto-debug]

Benchmark Results

Hardware: Apple M3 Ultra (28-core CPU, 96 GB RAM, 31.6 TFLOPS peak ANE dual-die)
Config: train_large --steps 30, synthetic 100K tokens

Default (ACCUM_STEPS=10)

Metric	main (baseline)	compile-io	Delta
Avg train	101.5 ms/step	100.7 ms/step	-0.8%
Wall time	16.2s	15.9s	-1.9%

With ANE_ACCUM_STEPS=20

Metric	main (baseline)	compile-io (accum=20)	Delta
Avg train	101.5 ms/step	98.5 ms/step	-3.0%
Wall time	16.2s	11.7s	-27.8%
Compile overhead	77.6%	71.7%	-5.9pp
exec() restarts	3	2	-33%

Findings

• Configurable ACCUM_STEPS is the main value: doubling from 10 to 20 cuts wall-clock by ~28% via fewer exec() restarts
• IOSurface pooling provides marginal benefit at this scale; larger benefit expected on longer training runs
• Env var tuning (ANE_ACCUM_STEPS, ANE_MAX_COMPILES) enables runtime adjustment without recompilation