Make the `_matmul` more generic

[GiggleLiu]

There is a Real type annotation added to matmul functions in recent versions.

Octavian.jl/src/matmul.jl

Lines 266 to 296 in 5ded095

	@inline function _matmul!(C::AbstractMatrix{T}, A, B, α, β, nthread, MKN) where {T<:Real}
	M, K, N = MKN === nothing ? matmul_sizes(C, A, B) : MKN
	if M * N == 0
	return
	elseif K == 0
	matmul_only_β!(C, β)
	return
	end
	W = pick_vector_width(T)
	pA = zstridedpointer(A); pB = zstridedpointer(B); pC = zstridedpointer(C);
	Cb = preserve_buffer(C); Ab = preserve_buffer(A); Bb = preserve_buffer(B);
	mᵣ, nᵣ = matmul_params(Val(T))
	GC.@preserve Cb Ab Bb begin
	if maybeinline(M, N, T, ArrayInterface.is_column_major(A)) # check MUST be compile-time resolvable
	inlineloopmul!(pC, pA, pB, One(), Zero(), M, K, N)
	return
	else
	(nᵣ ≥ N) && @goto LOOPMUL
	if (Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)
	(M*K*N < (StaticInt{4_096}() * W)) && @goto LOOPMUL
	else
	(M*K*N < (StaticInt{32_000}() * W)) && @goto LOOPMUL
	end
	__matmul!(pC, pA, pB, α, β, M, K, N, nthread)
	return
	@label LOOPMUL
	loopmul!(pC, pA, pB, α, β, M, K, N)
	return
	end
	end
	end

This function looks generic, we do not have to throw a type error on this function for non-real input. Now, I have to copy ~100 lines of code from Octavian to TropicalGEMM to fix the broken tests because the TropicalNumber is not a Real number but using exactly the same matrix multiplication code.

Actually, as @chriselrod mentioned before, any composite type with multiple fields is not supported yet, not because the type is not Real. May I open a PR to undo this change?

[chriselrod]

May I open a PR to undo this change?

Okay.