[GPU] Conv and Gemm fix

[dyoussif]

1. Fix data type for conv zp mask; set to s16 instead of s32. This is to avoid invalid data type combination for the mad instruction,mad dst:d src0:d src1:d src2:d.

Failing case

--mode-modifier=P --conv --engine=gpu --skip-impl=ref --allow-enum-tags-only=false --check-ref-impl=true -- dt=u8:s8:u8 --attr-scales=wei:per_oc --attr-zero-points=src:per_dim_1 --attr-post-ops=hardswish:0.271:0.314+linear:0.271:0.314 g240mb32ic240ih28oc240oh14kh3sh2ph0n"f191c263e53dbb3ce0c02a13f311a72a*1"

2. Fix gemm hang for the following case on Xe2:

--matmul --engine=gpu --dt=f16:s4:f16 --wtag=acb --attr-scales=wei:per_ocic:f16:128x1 --attr-zero-points=wei:per_ocic:u4:128x1 --attr-fpmath=f16:true --skip-impl=ref 3x96x512:3x512x64

Seems lookahead should match reqLoad

[dyoussif]

make test
disable device_cpu
enable device_gpu
disable benchdnn_all
enable benchdnn_matmul

[petercad]

@dyoussif, similar to the lines below, the lookahead should depend on whether we are repacking this data. If so -- yes, this patch is right. Otherwise, the original code is what's needed; with this patch we would load Aq/Bq too early if the group size exceeds the load chunk size for A/B.