[GPU] Tweak convolution plan

[echeresh]

Jira: MFDNN-14766

The original regression is due to the layout normalization changes. This resulted in some previously rejected configs to be enabled which caused the regression.

The PR includes two changes:

• Convolution plan tweak to prioritize the direct view optimization higher than SLM during retries
• Relaxed mad layout check for f16/bf16 types
  • In particular for "direct view" layouts we usually have overlapped blocks (same stride). Any of these two blocks are now checked during the vectorized dimension check

[echeresh]

make test perf-gpu
set primitive=conv deconv

[echeresh]

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_conv
enable benchdnn_deconv
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg
enable arch_gpu_xe3-lpg