Add FP requantize flow. Set float32 flow by default for llvm x86 targets with sse4.1 support.

[Icemist]

Added a new calculation_flow_type parament to the relay.qnn.op.requantize. This parameter is controlling the implementation flow of this function. Valid values: "int64", "float32", "float64".
The basic idea is that for some targets implementations other than "int64" (the only one at the moment) will be more productive.
Below some measurements were made on AMD Ryzen 7 5800H with TVM_NUM_THREADS=1

Performance with "llvm -mcpu=core-avx2" target:

Performance with "llvm" target:

Accuracy with "llvm -mcpu=core-avx2" target:

requantize flow	accuracy
UPWARD_float64	75,39%
UPWARD_float32	75,32%
UPWARD_int64	75,39%
TONEAREST_float64	75,39%
TONEAREST_float32	75,32%
TONEAREST_int64	75,39%

Accuracy with "llvm" target:

requantize flow	accuracy
UPWARD_float64	75,38%
UPWARD_float32	75,30%
UPWARD_int64	75,38%
TONEAREST_float64	75,38%
TONEAREST_float32	75,30%
TONEAREST_int64	75,38%

Additional changes:

• Added relay.qnn.op.requantize_config to use it at python "with" statement. This allows users to control the behavior of the requantize function not directly. It accepts two parameters: rounding and compute_dtype. It has a lower priority than these parameters passed directly to the requantize function. Note: compute_dtype will be "float32" for llvm x86 targets. For example:

mod, params = relay.frontend.from_pytorch(scripted_model, shape_list)
with tvm.transform.PassContext(opt_level=3):
	with relay.qnn.op.requantize_config(rounding="UPWARD", compute_dtype="float64"):
		lib = relay.build_module.build(mod, target=target, params=params)

• Added target_has_sse41 and target_is_x86 functions to tvm.topi.x86.utils python namespace
• Registered target_has.* functions from tvm.topi.x86.utils to call them from C++ code
• Added Floor, LogicalOr, Equal, Less and IsFinite relay operations in C++ tvm::relay namespace
• Added requantize_config validation tests

[Icemist]

Have you benchmarked on ARM? I think we should enable this only for x86 for now.
I turned it off for not x86. It is also now possible to manage the implementation of requantize that will be used.

[masahi]

cc @jwfromm

[masahi]

Looks good, only minor comments

[masahi]

Please go through your change and remove all uses of the term calculation flow

[masahi]

Do we need this level of details? I prefer dropping them. I don't think people would ever specify these targets...

I think sse4.1 - vnni are enough.

[Icemist]

I agree, sse4.1 looks good. Users can always use requantize_config to change the default behavior.
Done.

[masahi]

Very nice, just more minor comments and I'll merge this.

[masahi]

Please kick another CI job.

Addressed