fix: CpuGemmDirectConv2d: accumulate in fp32 unless in fast_mode

[Sqvid]

We have a bug in PyTorch+oneDNN+ACL where numerical errors were observed in certain f16 convs.
PyTorch issue: pytorch/pytorch#177245
oneDNN issue: uxlfoundation/oneDNN#5106

The root cause is that we are accumulating in f16 rather than in f32 even when fast_mode is false. The reason this happens is because CpuGemmDirectConv2d::configure() calls CpuGemmAssemblyDispatch::configure() here:

ComputeLibrary/src/cpu/operators/CpuGemmDirectConv2d.cpp

Line 136 in d619e50

_gemm_asm_func->configure(src, &_perm_weights, biases, dst, asm_info);

Which does the following:

ComputeLibrary/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp

Lines 853 to 856 in d619e50

	// If fast_mode is disabled, we must enable it when fp32 accumulation is not set for fp16.
	bool is_fp16 =
	a->data_type() == DataType::F16 && b->data_type() == DataType::F16 && d->data_type() == DataType::F16;
	bool fast_mode = info.fast_mode || (is_fp16 && !info.use_fp32_acc);

My solution is therefore to change CpuGemmDirectConv2d::init_assembly_metadata() such that use_fp32_acc is true unless enable_fast_mode is set. This resolves the bug at the oneDNN level. Hopefully @robert-hardwick can confirm if this fixes the PyTorch bug as well.

[Dongsung-arm]

Looks good to me.
This matches the intended behavior: FP32 accumulation by default, and only disabled when enable_fast_math is set.