perf(rms_norm): use fused reduce_l2_norm path (~48× faster)

[sbryngelson]

Summary

rms_norm lowered to a reshape to [M,D,1,1] followed by reduce_sum over the
channel axis. That reduction falls off the ANE's fast reduction tile past
~256 rows, so rms_norm ran 16–27× slower than layer_norm — even though
RMS-norm is structurally cheaper (one reduction, no centering) — and scaled
super-linearly with row count.

Measured on M5 (H17s), gamma=1:

shape	before	after	speedup
256×256	160 µs	~110 µs	1.5×
512×512	1791 µs	~110 µs	16×
1024×1024	6882 µs	168 µs	41×
2048×1024	13665 µs	257 µs	53×

Fix

Re-lower through the same fused reduce_l2_norm over the last axis that
l2_norm already uses, since the two are mathematically identical:

rms(x) = x / sqrt(mean(x²)+eps) · g  =  x · √D / sqrt(sum(x²)) · g     (reduce_l2_norm = sqrt(sum x²))

eps becomes a safe-divide floor on the norm (same pattern as l2_norm), and
the √D rescale is folded into the gamma weight, so the emitted op count drops
as well (no [M,D,1,1] reshape, no separate square/reduce_sum/scale/rsqrt chain).

Correctness / tests

• Max error vs fp32 reference RMS ≤ 0.007 across shapes (existing
  test_nn_blocks::rms_norm_linear_silu tolerance is 0.03).
• tests/test_builder_guards.py (2D guard) unchanged and passing.
• Full suite: 527 passed on M5/H17s.

RMS-norm is ubiquitous in modern transformers (LLaMA/Qwen/etc.), so this is a
high-impact, low-risk lowering fix.