perf(PULPOpen): add a vector bias without materialising the broadcast

[runwangdl]

AddLayer.computeShapes rewrites the lower-rank operand to the higher-rank one's shape. A Linear exported as MatMul + Add hands the Add a bias of shape [O] against an activation of [1, M, O], so the bias ends up stored as M identical rows — on CCT-2 at 64 tokens, 32 KB to carry 512 B, and the eight live at the peak account for 252 KB.

What changes

Keep the bias at [O] and teach the three places that assumed equal shapes:

PULPFloatAddTemplate	walks the second operand with a wrapped row index — advanced and reset, not taken modulo, since there is no hardware divide on this core
PULPBroadcastAddTileConstraint (new)	constrains only the trailing axis, and gives the broadcast operand the trailing slice of the output cube instead of the whole cube
PULPAddLayer	skips the rewrite when the smaller operand has a single non-unit axis

Each falls back to the inherited elementwise path when the operands do have equal shapes, so an ordinary Add is untouched. Targets/Generic is not modified.

Measured

GAP9, --l1 122000 --defaultMemLevel L3, L3 peak, address-deduplicated:

model	before	after
CCT linear probing	1954.9 KB	1702.6 KB	−12.9%
CCT LoRA on one block	1965.5 KB	1713.0 KB	−12.8%
CCT full FT on one block	2376.8 KB	2252.0 KB	−5.3%

The bias's own L3 allocation drops from cl_ram_malloc(32768) to cl_ram_malloc(512), and the activation side is unchanged at 512.0 KB — the expected shape of the saving, since a bias is persistent rather than an activation.

Verification, including what does not pass

• linear probing — Errors: 0, 37.68M train cycles
• one-block full fine-tuning — Errors: 0, 54.94M train cycles
• one-block LoRA — does not build, with or without this change. hoistConstant asserts one consumer per Constant, and a LoRA graph's frozen transposed weight feeds both the forward Gemm and its gradient. A control run on this branch with this commit removed fails with the identical assertion, so the limitation is pre-existing. Its figure in the table therefore comes from the allocation plan, not from a run.

Stacked on #50, which adds the GAP9 registration this builds on.

🤖 Generated with Claude Code

[runwangdl]

Folded into #54 so this ships as a single merge. The branch is kept, and the content is in #54 with its description.