Rework LTO, Intel Assembly syntax and default compile-options

[mratsim]

This tries to address #303 #295 codex-storage/nim-codex#625 (comment) following making Constantine compatible with LTO #231

• Intel assembly is used on a per-file basis to avoid interfering with other libraries like secp256k1 doesn't work
• On Apple MacOS/iOS it will need to be explicitly enabled, due to:
  • Clang not supporting Intel syntax:
    • This has not be upstreamed in Apple Clang: https://reviews.llvm.org/D113707
    • Opened issue to Apple: FB12137688
  • Intel Apple CPUs being outdated and not being used for high-performance cryptography and Intel syntax being irrelevant for ARM CPUs.

[mratsim]

MacOS is so annoying ....

• Apple Clang without LTO, removes some global variables (ADX) when compiling to a static library
• LLVM Clang v15 with -flto=thin / -d:lto_incremental does not find enough registers .... but it does with full LTO
• And Apple Clang not supporting Intel syntax

[mratsim]

@dryajov @benbierens @markspanbroek

Constantine now uses Intel Syntax only if LTO is chosen instead of always, hence it isn't needed to set it in libraries that depend on it.

constantine/constantine/platforms/config.nim

Lines 31 to 56 in 747830b

	const UseAsmSyntaxIntel* {.booldefine.} = defined(lto) or defined(lto_incremental)
	## When using LTO with AT&T syntax
	## - GCC will give spurious "Warning: missing operand; zero assumed" with AT&T syntax
	## - Clang will wrongly combine memory offset and constant propagated address of constants
	##
	## Intel syntax does not have such limitation.
	## However
	## - It is not supported on Apple Clang due to missing
	## commit: https://github.com/llvm/llvm-project/commit/ae98182cf7341181e4aa815c372a072dec82779f
	## Revision: https://reviews.llvm.org/D113707
	## Apple bug: FB12137688
	## - Global "-masm=intel" composition with other libraries that use AT&T inline assembly
	##
	## As a workaround:
	## - On MacOS/iOS upstream Clang can be used instead of Apple fork.
	## - Do not use LTO or build Constantine as a separate library
	##
	## Regarding -masm=intel:
	## - It might be possible to use Intel assembly is used on a per-file basis
	## so that we do not affect other libraries that might be compiled together with Constantine.
	## Generating an object file works but the final linking step assumes AT&T syntax and fails.
	## - Surrounding code with ".intel_syntax noprefix" and "att_syntax prefix"
	## doesn't work with memory operands.
	when UseAsmSyntaxIntel:
	{.passC: "-masm=intel".}

If using LTO:

• Intel syntax is necessary for Clang, recommended for GCC (or it will spam spurious warnings that may denial-of-service the console) and so activated by default.
• Dynamic linking solves assembly syntax incompatibility with other libraries like libsecp256k1
• On MacOS, Intel Syntax requires upstream LLVM Clang, not Apple Clang
• If static linking, using the same compiler and the same major version is strongly recommended