mha: size the K/V arg specs by num_KV_heads, not num_heads

[atassis]

design.py declares Q and O as (heads, S_q_pad, d) but K and V as
(num_KV_heads, S_kv_pad * d). get_arg_spec() sized all four from num_heads, so
under GQA the K/V specs are over by num_heads / num_KV_heads. On the GQA param
already in test.py (seq_len=16384, d=64, heads=8, pipelines=8, kv_heads=2) that is
8388608 against 2097152, a factor of 4.

Two things kept it hidden. num_kv_heads=0 normalizes to heads in design.py, which
makes the old formula accidentally correct for standard MHA, and run_test reads only
spec.direction, taking input buffers from the supplied tensor rather than from
spec.shape. So nothing consults the wrong value today. It surfaces inside an
OperatorSequence, where compilation/sequence.py asserts the MLIR arg count equals
the one computed here.

Added

• iron/operators/mha/test.py: test_arg_spec_matches_design_shapes, covering GQA and
  standard MHA. It asserts against the shapes design.py declares rather than against
  literals, so it fails if either side drifts. Device-free, and deliberately not reusing
  get_params() -- the GQA case there is marked extensive, which is half of why this
  went unnoticed.

Changed

• iron/operators/mha/op.py: get_arg_spec() sizes Q/O from num_heads and K/V from
  num_KV_heads, applying the same 0 -> num_heads normalization design.py does.

Removed

Known limitation

_calculate_seq_padding hardcodes the block size to 64, while design.py derives
S_q_pad and S_kv_pad from B_q and B_kv separately. That is correct today only
because both default to 64 and op.py never overrides them, and __post_init__ sets
self.B_q/self.B_kv which the helper ignores. Left alone here to keep this to one
defect.