metal: PQ2_0 and Q2_0 mat-vec dot in exact float (+5-7% decode)

[bri-prism]

The base-4 analogue of the exact-float extraction from #157, applied to the two 2-bit mat-vecs. PQ2_0 and Q2_0 share one codec at group sizes 128 and 64, so one templated dot serves both.

What

	decode before	decode after
PQ2_0, smaller dense	285.8	305.1 tok/s, +6.8%
PQ2_0, larger dense	80.2	83.1 tok/s, +3.6%
Q2_0, smaller dense	283.6	297.3 tok/s, +4.8%

M5 Pro, llama-bench -p 512 -n 128 -r 3, two interleaved passes. Prefill unchanged in all cases; the mat-mul path does not use this routine.

Why

The dot accumulated each 2-bit field through two ands, two bool tests, two selects and two adds per element, none of which this ISA can co-issue with the floating-point adds.

How

A byte is a base-4 fraction of 256. With u = b/256 and g_k = floor(4^k * u), exact in fp32, the floor chain peels the fields from the top down, because plain bit packing keeps field 0 in the low bits: g_1 is field 3, g_2 - 4*g_1 is field 2, g_3 - 4*g_2 is field 1 and b - 4*g_3 is field 0. Summing field times activation over a byte collapses to g_1*(y_3 - 4*y_2) + g_2*(y_2 - 4*y_1) + g_3*(y_1 - 4*y_0) + b*y_0, whose coefficients depend only on the activations, so the caller stages those in place of the raw y and reuses them across every row. Three floors and four fmas per byte, no integer work. The shared select-chain overload is left in place for its other callers.

The field order was checked against the packing over all 256 bytes before the kernel was touched. A first derivation that assumed the base-3 digit order was wrong and was caught by that check; the base-3 codec stores digit 0 as the leading fractional digit, which is why the same chain runs bottom-up there and top-down here.

What was tried and rejected

The same rewrite on Q1_0 is a measured regression of 5 to 7 percent on two dense models and is not included. With one bit per element the select is already a single op, so seven floors per byte cannot beat it and floor's lower throughput loses. The rule that falls out: the float-floor form pays only where the integer path costs two or more selects, or a multiply-shift chain, per element.

Where the 2-bit formats now sit

Against a weight-bandwidth ceiling of roughly 270 GB/s, PQ2_0 decode on the larger model is at about 75 percent of the ceiling after this change, so there is little ALU headroom left there. The base-3 formats sit lower (PTQ1_0 about 61 percent, TQ1_0 about 49 percent) and remain the place where kernel work still moves the number.

Correctness

test-backend-ops on Metal, rebuilt on this branch's base: MUL_MAT 45 cases pass for each of Q2_0 and PQ2_0 with exit code 0 out of 1719 total; MUL_MAT_ID 75 pass for each with exit code 0 out of 1022; Q1_0 is unchanged and passes the same. The collapse is exact; results are not bit-identical to the select chain because the coefficient staging reassociates the fp32 sum. A/B arms were separate build directories with the embedded shader confirmed distinct before each measurement.