ggml : add broadcast support for BLAS ggml_mul_mat()

[ggerganov]

For F16 and F32 CPU matrix multiplications BLAS will likely remain necessary, especially on Apple Silicon where we have Accelerate CBLAS using the AMX coprocessor. So we need to resolve the following TODOs:

ggml/src/ggml.c

Lines 10796 to 10799 in 08c57df

	// TODO: handle case when src0 is broadcast-able into src1 across 2nd,3rd dimension
	// ref: https://github.com/ggerganov/ggml/pull/224
	GGML_ASSERT(ne02 == ne12);
	GGML_ASSERT(ne03 == ne13);

ggml/src/ggml.c

Lines 10782 to 10785 in 08c57df

	// TODO: handle case when src0 is broadcast-able into src1 across 2nd,3rd dimension
	// ref: https://github.com/ggerganov/ggml/pull/224
	GGML_ASSERT(ne02 == ne12);
	GGML_ASSERT(ne03 == ne13);

This will fix SAM inference on Apple Silicon: #459

The broadcast logic is implemented in the same function for the non-BLAS branch:

ggml/src/ggml.c

Lines 10905 to 10909 in 08c57df

	// broadcast factors
	const int64_t r2 = ne12/ne02;
	const int64_t r3 = ne13/ne03;

Should be simple enough to adapt, but need to make sure it works correctly. Probably the best way to test it is to use the SAM inference example, which works without BLAS but asserts with BLAS.