[SYCL][bfloat16] Simplify bfloat16 class

[bader]

This change makes a few changes aiming to make the code more readable.

• Move declaration of the bfloat16 class right on top of the header file
  and move all declarations with the implementation helpers after the
  class declaration.

• Converted ConvertToBfloat16 class to namespace. The class doesn't
  represent any object, it's a collection of static methods.

• Declare public API of the class first.

• Move Bfloat16StorageT declaration from the detail namespace directly
  into bfloat16 class definition.

• Co-locate declaration of functions with similar functionality (e.g.
  conversion functions).

• Simplified uniry operator- implementation by removing additional
  branch for the SPIR target. Generic code for SPIR target produces
  exactly the same code.

• Added compile time checks for the array size inside conversion helper
  functions. Device side implementation of these functions work
  correctly only for certain array sizes.

• Include C++ header instead of C (i.e. cstdint instead of stdint.h).

• Applied "no-else-after-return" and "no-braces-for-single-statement-if"
  coding style changes.

[cperkinsintel]

Nice.

When I saw the title I was hoping that it would also give blfloat16 a constexpr contstructor and operator==, but still a bridge too far.

[bader]

When I saw the title I was hoping that it would also give blfloat16 a constexpr contstructor and operator==, but still a bridge too far.

That's good feedback. I suggest we apply a "rule of zero" to this class to make constructor constexpr.

Regarding operator==. I'll try to make all binary operators constexpr.

[bader]

When I saw the title I was hoping that it would also give blfloat16 a constexpr contstructor and operator==, but still a bridge too far.

That's good feedback. I suggest we apply a "rule of zero" to this class to make constructor constexpr.

Regarding operator==. I'll try to make all binary operators constexpr.

Okay, I think I understand why both these ideas haven't been implemented yet.

1. We can't apply "rule of zero" because we have defined conversion constructors from float and sycl::half data types.
2. We can't make operator== constexpr because SPIR-V implementation uses __devicelib_* functions, which aren't constexpr.

Let's leave constexpr improvements for future patches.

I suppose we significantly simplify the implementation by switching to from uint16_t to __bf16 data type for value field. Unfortunately, __bf16 is currently supported only by just a few targets. x86 is the only target which we regularly test, and which supports __bf16 data type.

[bader]

Pre-commit failures are known issues tracked by #17251, #17026.

One failure on NVIDIA GPU hasn't been reported before, so I opened #17285. It doesn't seem to be related to my change as this test passed for 3adc510 commit, which was the PR state before I moved one line with the code comments.