perf: batch_size 16384 → 8192 — 2x faster GEMMs, more frequent updates

At 16384, each GEMM saturates 97% of 132 SMs — no room for multi-stream
parallelism. At 8192, GEMMs use ~50% SMs, allowing branch streams and
aux parallelism to fill idle SMs. 2x more steps/epoch but each ~2x faster.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

config/gpu/h100.toml

@@ -1,6 +1,6 @@
 # H100 PCIe/SXM (80GB VRAM) -- full production
 [training]
-batch_size = 16384
+batch_size = 8192
 num_atoms = 52
 buffer_size = 500000
 hidden_dim_base = 256