[SYCL] Speed up device compilation for half and bfloat16

[aelovikov-intel]

• std::istream/std::ostream aren't usable on device, so limit operator<</operator>> to declaration only to limit device includes to <iosfwd> instead of much heavier <istream>/<ostream>.
• Use "lighter" <optional> to get std::hash

[aelovikov-intel]

half's hidden friends already do this, no idea why ESIMD needed this in the first place.

[sarnex]

i think we had an implementation of half in esimd before sycl::half was implemented, probably we forgot to remove it

[bader]

i think we had an implementation of half in esimd before sycl::half was implemented, probably we forgot to remove it

I see stream operators implemented before ESIMD.

ESIMD implementation added by: f34ba2c

Here is sycl::half implementation for the same SHA:

llvm/sycl/include/CL/sycl/half_type.hpp

Lines 518 to 528 in f34ba2c

	inline std::ostream &operator<<(std::ostream &O, cl::sycl::half const &rhs) {
	O << static_cast<float>(rhs);
	return O;
	}
	inline std::istream &operator>>(std::istream &I, cl::sycl::half &rhs) {
	float ValFloat = 0.0f;
	I >> ValFloat;
	rhs = ValFloat;
	return I;
	}

[sarnex]

esimd doesn't work on the host now, im not sure about at the time, but there is no point in having this code for the host

[sarnex]

i think we had an implementation of half in esimd before sycl::half was implemented, probably we forgot to remove it