1. Add a new API -- sbgemv to support bfloat16 based gemv
2. Implement a generic kernel for sbgemv
3. Implement an avx512-bf16 based kernel for sbgemv

Signed-off-by: Chen, Guobing <guobing.chen@intel.com>

...not a matrix size. For GEMM_MULTITHREAD_THRESHOLD=4
(the default value) this does not change anything but
for other values it make the GEMM and GEMV thresholds
changing in the same way.

Close #742

Ref #727

Close #722

Thank matzeri's patch.

Ref #478

For gemv_t, directly use malloc to create the buffer.

* stack_alloc_size is needed after the implementation call
but it may be overwritten if it's optimized to a register,
because some gemv implementation (ex: dgemv_n.S) do not
restore all register (ex: r10).
* do the same in ger.c for the same reasons even if the bug
has not been observed.

ger and gemv call blas_memory_alloc/free which in their turn
call blas_lock. blas_lock create thread contention when matrices
are small and the number of thread is high enough. We avoid
call blas_memory_alloc by replacing it with stack allocation.
This can be enabled with:
make -DMAX_STACK_ALLOC=2048
The given size (in byte) must be high enough to avoid thread contention
and small enough to avoid stack overflow.

Fix #478

Signed-off-by: Timothy Gu <timothygu99@gmail.com>