rms_norm: fix bf16 accuracy (f32 normalize + conv_even) and parameterize epsilon

[atassis]

Summary

Two related changes to the aie2/aie2p bf16 RMSNorm kernel and its IRON operator:

1. An accuracy fix: normalize in f32 and round once, and set the rounding mode to
   round-to-nearest-even (conv_even).
2. Make epsilon a parameter instead of a hardcoded 1e-5f.

The default behavior is unchanged (epsilon still defaults to 1e-5). Callers opt
into a different value with RMSNorm(..., epsilon=...).

Motivation

I hit this bringing a small LLM decode (Gemma-3-270M, 18 fused layers) up on
aie2p. The generated tokens matched a host reference for the first few decode
steps and then flipped on a step where two candidate tokens were close in logit
space. The graph itself was correct: an f32 (and a faithful round-to-nearest
bf16) reference of the same graph reproduced the reference output. The device was
noisier than a correct bf16 forward should be, and the error grew roughly
linearly with depth rather than as sqrt(depth).

Two things in the RMSNorm kernel caused that:

• inv_rms (the reciprocal RMS scalar) was cast to bf16 and then multiplied
  bf16*bf16, so the entire normalized vector was scaled by one imprecise scalar
  per norm. That is a coherent, same-direction error per norm. Over a deep
  residual stream it accumulates roughly linearly, which is what made it large
  enough to flip a near-tie argmax; unbiased round-to-nearest error would grow
  only as sqrt(N) and partly cancel.
• The kernel set no rounding mode, so it ran under whatever a previously
  dispatched kernel left in the core's global rounding state. set_rounding is
  global per core, so in a fused graph a kernel that leaves floor/truncation set
  poisons the ones that follow.

Separately, epsilon was hardcoded to 1e-5f (Llama's value). Gemma uses 1e-6,
and it is not a rounding detail there: Gemma's deep post-feedforward sandwich
norms operate on near-zero inputs where mean(x^2) is comparable to epsilon, so
epsilon sets the normalization scale.

What changed

Commit 1, "normalize in f32 and set conv_even rounding":

• aie_kernels/{aie2,aie2p}/rms_norm.cc: normalize through an accfloat
  accumulator (mirroring layer_norm.cc) and round once at the output; call
  set_rounding(conv_even) at both entry points.

Commit 2, "parameterize epsilon":

• aie_kernels/{aie2,aie2p}/rms_norm.cc: rms_norm_general /
  rms_norm_bf16_vector / weighted_rms_norm take a float epsilon argument.
• iron/operators/rms_norm/{design,design_weighted}.py: declare the np.float32
  kernel argument and pass epsilon (a per-instance compile-time constant, the
  same mechanism the axpy op uses for its scalar factor).
• iron/operators/rms_norm/op.py: add an epsilon field (default 1e-5).
• iron/operators/rms_norm/reference.py: accept eps and align the formula to
  the kernel, 1/sqrt(mean(x^2) + eps) rather than 1/(sqrt(mean(x^2)) + eps).

Testing

On aie2p (Strix):

• Isolated weighted RMSNorm rel-L2 vs a host reference improved from ~0.008 to
  ~0.0025.
• A fused 18-layer Gemma-3-270M decode that was previously flipping a token now
  matches its host reference exactly (8 of 8 greedy tokens), deterministic across
  repeated runs, with the op passing epsilon=1e-6.

The aie2 (Phoenix) kernel gets the same source change and compiles; I do not have
aie2 hardware, so it is compile-verified, not silicon-verified. The change mirrors
the aie2p one and the existing aie2 layer_norm.cc accfloat path.

[andrej]

Thanks for the accuracy improvements! Cleaned up the comments a bit (sorry for being anal), if this passes ready to merge, feel free to remind me if I forget again with everything else going on.

[atassis]

Fixed CI

[atassis]

Oh, and yes, no issues, they are all reasonable:)
I'll admit why I don't clean up everything by myself- sometimes even after proof reading the code and comments I tend to internally approve smth even after AI review and else. So it is easier for me to suggest smth, get a review and then fix that.
I do that because I find it easier for someone to have a fast look and direct me, instead of talking more time to try and deduce myselves.
For example, after talking to Joseph I have found another idea, which led to me through few more bugs, which you have seen patched already. And that idea is not fully researched yet