Fix A2 fast-path prewarm count to match the generic WaveNet

[rhaist]

Summary

A2FastModel (the A2 fast-path WaveNet, a drop-in for the generic WaveNet)
computed its prewarm sample count as the layer-stack lookback distance —
Σ (kernel_size-1)·dilation + (head_kernel-1) — instead of the receptive
field, which is one greater. The generic path seeds mPrewarmSamples at 1
(the sample being produced) before adding the same lookback terms
(model.cpp), so the fast path warmed up by one fewer sample than the
model it replaces.

Impact

prewarm() runs process() in whole maxBufferSize blocks until the count is
reached, so the off-by-one only changes the block count when the total (6346
for the A2 shape) is an exact multiple of the buffer size. Power-of-2 buffers
(64/256/4096) mask it — which is exactly why the existing equivalence test
(test_matches_generic, blocks 64/256) never caught it. On a buffer size that
divides 6346 (e.g. 334), the two paths warm up by a different number of
blocks and their first post-Reset output diverges.

Low severity (no audible artifact at typical buffer sizes, no RT-safety
impact), but a real contract violation for a path whose whole purpose is to be
numerically identical to the generic WaveNet.

Fix

• Seed prewarm at 1 to match the generic receptive-field formula, with a
  comment explaining the anchor so the count isn't re-derived incorrectly again.
• Add test_prewarm_matches_generic_{nano,standard}: builds both the fast and
  generic model from the same config and asserts equal GetPrewarmSamples().

Verification

Confirmed the test catches the regression — reverting the seed to 0 aborts
the new assertion; with the fix the full run_tests suite passes (A2 fast path
is built by default via NAM_ENABLE_A2_FAST).

[sdatkinson]

Thanks again!