metal : add TQ2_0 support

[ggerganov]

Overview

Implement missing TQ2_0 ops in the Metal backend.

Hoisting the per-byte coefficients

The TQ2_0 mul_mv kernel packs 4 quantized values per byte v, stored as 2-bit fields

$$f_j = (v \gg 2j) \,\&\, 3, \qquad j = 0..3$$

A byte's dot-product contribution is $\sum_j (f_j - 1)y_j$. Each field is a difference of consecutive floor-divisions by powers of 4:

$$f_j = \left\lfloor v / 4^{j} \right\rfloor - 4 \left\lfloor v / 4^{j+1} \right\rfloor$$

Substituting all four fields and collecting terms, the interior floors cancel (telescoping), leaving only the boundary floors — the byte divided by successive powers of 4:

$$\sum_j (f_j - 1)\,y_j \;=\; y_0 v + (y_1 - 4y_0)\left\lfloor \tfrac{v}{4} \right\rfloor + (y_2 - 4y_1)\left\lfloor \tfrac{v}{16} \right\rfloor + (y_3 - 4y_2)\left\lfloor \tfrac{v}{64} \right\rfloor - (y_0 + y_1 + y_2 + y_3)$$

Hence the inner loop's four f values are exactly $v$, $\lfloor v/4\rfloor$, $\lfloor v/16\rfloor$, $\lfloor v/64\rfloor$ — the survivors of the bit-field decomposition.

The coefficients $y_0$, $y_1-4y_0$, $y_2-4y_1$, $y_3-4y_2$ and the constant $y_0+y_1+y_2+y_3$ depend only on y, which is the same for every row. So the kernel "hoists" them out of the row loop, building them once per block and reusing them for all nr0 rows:

// once per block:  coef[j] = (Y0[j], Y1[j]-4Y0[j], Y2[j]-4Y1[j], Y3[j]-4Y2[j])
// once per block:  sumy    = (Y0[j]+Y1[j]) + (Y2[j]+Y3[j])
// per row:         sum = -sumy + coef[j][0]*v + coef[j][1]*f1 + coef[j][2]*f2 + coef[j][3]*f3

Net effect: the y-dependent setup runs once per block instead of once per (block × row), removing work proportional to the row count — and the per-byte work becomes pure float ops (floor + multiply) instead of integer bit masking/shifting.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. llama.cpp:DeepSeek-v4-Flash-0731