ARM64 SME: fix zero-scalar handling and clean up direct kernels

[moluopro]

1. Fix zero alpha/beta semantics

The concrete issues are:

• By BLAS semantics, the input matrix C must not be referenced when beta == 0.
• When alpha == 0 or K == 0, only the beta*C update is required; A and B should not be read.
• The current SME direct path bypasses the generic Level-3 handling, so NaNs in C or A/B can pollute the output through 0 * NaN.

2. Use integer round-up for padded SME dimensions

ceil((double)M / (double)vl_elms) * vl_elms

is replaced with integer arithmetic:

(((uint64_t)M + vl_elms - 1) / vl_elms) * vl_elms

3. Clean up const and unused warnings

• Change the read-only source pointer in the SSYMM preprocess helper to const float *restrict.
• For SSYMM, compile only the preprocess helper actually used by the current UPPER / LOWER object variant.
• Change the read-only B parameter in SSYRK kernel_2x2() to const float *.
• Change the read-only B_T and A_T parameters in SSYR2K kernel_2x2() to const float *.
• Compile sve_cntw() and ldb only in the TRANSA / !TRANSA variants that actually use them.