cpu: x64: conv: remove unnecessary work around

[yair-obodovsky]

Remove unnecessary flow that touches every page in the input buffer in conv layer.
Related to jira: https://jira.devtools.intel.com/browse/MFDNN-14582

[yair-obodovsky]

make test
set test_scope=NIGHTLY

[vpirogov]

There's one more spot with similar workaround:

oneDNN/src/cpu/x64/jit_brgemm_conv_bwd_strided.cpp

Lines 783 to 789 in 3e50d90

	if (is_amx && inp_buffer) {
	// Workaround: for some machines SEGFAULT possible on tile load
	// if the page was not touched before it
	for (dim_t i = 0; i < jcp.inp_buffer_size;
	i += brgemm_convolution_bwd_utils::P4K)
	inp_buffer[i] = 0;
	}

I do not see anything in matmul, but it's worth double checking.

[yair-obodovsky]

There's one more spot with similar workaround:

oneDNN/src/cpu/x64/jit_brgemm_conv_bwd_strided.cpp

Lines 783 to 789 in 3e50d90

	if (is_amx && inp_buffer) {
	// Workaround: for some machines SEGFAULT possible on tile load
	// if the page was not touched before it
	for (dim_t i = 0; i < jcp.inp_buffer_size;
	i += brgemm_convolution_bwd_utils::P4K)
	inp_buffer[i] = 0;
	}

I do not see anything in matmul, but it's worth double checking.

Removed from conv_bwd. We didn’t find it anywhere else.