Enable batch_size>1 native decode (incl. true sub-64); bump compiler + tools

[booth-algo]

Summary

Enables batch_size>1 native decode end-to-end, including true sub-64
(mlen 16/32 < head_dim 64). Pairs the PLENA_Compiler decoder/kernel fix with the
PLENA_Tools comparison reader.

• emulator_runner.py: forward rows_per_batch/active_seq_per_batch from
  comparison_params to compare_vram_with_golden, so a batch>1 output compares
  only the active (rpb-strided) rows rather than all physical rows (which include
  tile-align padding and false-negatived the match rate, e.g. 29% on a ~99%-correct run).
• Bump PLENA_Compiler → feat/multibatch-decoder (batch>1 MHA decoder + head_dim>mlen kernel).
• Bump PLENA_Tools → feat/multibatch-comparison (rpb-strided active-row reader).

Companion PRs

• Support batch_size>1 in native MHA decoder, incl. true sub-64 (head_dim>mlen) PLENA_Compiler#55
• verification: extract rpb-strided active rows for batch>1 comparison PLENA_Tools#6

Submodule pointers reference those PR branches; they'll be re-bumped to the merged
main commits once the companions land.

Test plan (numerical, allclose PASS)

• batch=2 @64/64/16 seq=4 = 97.37% (was a 29% false-negative before the harness fix)
• batch=2 @16/16/4 seq=4 = 99.07%, @32/32/4 = 99.67% (sub-64, no NaN)
• batch=1 @64/64/16 seq=4 = 99.96% (no regression)