Add loops for ATen compiler MHA asm gen to reduce instruction count

[booth-algo]

Summary

Loops the ATen per-head MHA attention helper emission to reduce instruction memory pressure while preserving the old Python-unrolled path.

Changed helpers:

• _online_softmax_asm
• _scale_o_asm
• _final_scaling_asm
• _pv_multiply_asm
• _reset_fpsram_asm
• _reset_vram_asm

Default emission now uses hardware loops. The existing public unroll path is preserved: ATEN_UNROLL=1 / unroll_loops=True now also sets unroll_attention=True, so the attention helpers use the Python-unrolled path. Tests and harnesses can still override prog.unroll_attention directly after construction for A/B comparisons.

CLM-60M Native Layer 0 Counts

Rerun locally from the simulator workspace with compile_hf_model(model, seq_len=64, hidden_size=None, inter_dim=None, num_layers=1). Native dims: hidden=384, inter=1408, heads=6, kv_heads=2, head_dim=64.

Metric	Previous	Current	Change
Total ASM source lines	35,479	15,367	-20,112 (-56.7%)
Actual static instruction lines	34,041	14,403	-19,638 (-57.7%, 2.36x smaller)
Comment / metadata lines	1,438	964	-474 (-33.0%)
Loop-expanded dynamic instructions	645,334	649,762	+4,428 (+0.69%)
Estimated cycles	8,915,366	8,919,794	+4,428 (+0.05%)
Estimated ms @ 1GHz	8.915366	8.919794	+0.004428 (+0.05%)
C_LOOP_START static lines	248	296	+48

Verification

Companion simulator branch asm-count-verification adds the harnesses and report. Results from that branch:

• ATen MHA seq=64, head_dim=64: static instructions drop from 2,960 to 128; estimated cycles increase about 3%.
• ATen MHA seq=64, head_dim=128: static instructions drop from 4,399 to 169; estimated cycles increase about 3%.
• Transactional emulator golden check for looped ATen MHA passed against PyTorch SDPA with 100% allclose pass rate under repo thresholds.
• Additional flag check: ATEN_UNROLL=1 constructs PlenaCompiler with unroll_loops=True and unroll_attention=True; ATEN_UNROLL=0 leaves both false.