Add NEON for av2_update_nbr_diagonal

[jjustiss-apple]

The av2_update_nbr_diagonal function updates context and magnitude arrays
along the diagonal scan path in trellis coded quantization. It accounts
for roughly 5% of ARM encode time at cpu-used=1.

This adds a NEON implementation that vectorizes across the 8 trellis
states using 64-bit vectors. Key techniques: vtbl1 for state-chain
table lookups, vrhadd for rounding halving adds, vsli for bit-field
packing, and an identity fast-path to skip redundant table lookups on
the common path.

Also deduplicates kTcqBaseMaxTbl from separate static definitions in the
C and AVX2 files into a single shared definition in trellis_quant.h.

Micro-benchmark results (Apple M-series, P-core):

Kernel	C (ns)	NEON (ns)	Speedup
update_nbr_diagonal 4x4	130.7	9.2	14.2x
update_nbr_diagonal 8x8	144.5	10.6	13.6x
update_nbr_diagonal 16x16	204.4	12.7	16.0x
update_nbr_diagonal 32x32	340.5	20.0	17.0x

CTC Results (RA, cpu-used=1, 33 frames, ARM):

Metric	Delta
Encode time	-4.9%
BD-rate Y	0.000%
BD-rate Cb	0.000%
BD-rate Cr	0.000%

[jjustiss-apple]

@yunqingwang1 I don't want to burden you with too many review requests. Whom else would you suggest that might be willing to help review ARM NEON PRs?

[yunqingwang1]

@yunqingwang1 I don't want to burden you with too many review requests. Whom else would you suggest that might be willing to help review ARM NEON PRs?

@jjustiss-apple Please feel free to add @jianj-g as a reviewer. He is quite experienced with SIMD optimization.

[jjustiss-apple]

Thank you @jianj-g for the careful review, I believe I've addressed everything.

[urvangjoshi]

@jianj-g : any other comments?

[urvangjoshi]

Pls rebase and force-push (don't create merge commit) @jjustiss-apple