Make bignum inline assembly replacement more flexible

[hanno-becker]

Suggested enhancement

Currently, we only allow inline assembly for single long-multiply-accumulate steps within scalar*bignum long multiplications (MULADDC_CORE). There are at least two ways this should be made more flexible to allow for higher performance bignum implementations:

• (Simple) Allow the optimization of multiple consecutive occurrences of MULADDC_CORE, as discussed in Improve inline assembly for Cortex-M + DSP #5360. Ultimately, allow replacing the entire mpi_mul_hlp() routine.
• (More complex) Allow a larger step/window size during Montgomery multiplication: Instead of iterating through inputs in steps of mbedtls_mpi_uint, allow iteration in steps of mbedtls_mpi_uint[N]. This will allow to significantly reduce the number of memory operations used multiplications, esp. on the A-profile where we have 31 GPRs of which the current MULADDC_CORE uses only very few.

At first I wondered whether the second point should be implemented by changing the typedef's for mbedtls_mpi_uint, but that's an API break since mbedtls_mpi_uint is part of mbedtls_mpi_mul_int(). It seems better to directly modify the granularity of the two loops around mpi_mul_hlp() we have in the code: Firstly, mbedtls_mpi_mul_mpi() (https://github.com/ARMmbed/mbedtls/blob/development/library/bignum.c#L1487-1488), and secondly, mpi_montmul() (https://github.com/ARMmbed/mbedtls/blob/development/library/bignum.c#L1911-L1935).