[sw,crypto] Replace Barrett reduction with Solinas in P-384 lib.

Uses two Solinas reduction implementations, a specialized one for
reduction modulo p and a slower, more generic one for reduction modulo
n.

After this change, there is substantial speedup for P-384 sign/verify
and not much change in code size:
- 1.27x speedup for p384_ecdsa_sign_test (from 0x1c3ec6 cycles to
  0x164200 cycles)
- 1.26x speedup for p384_ecdsa_verify_test (from 0x142f77 cycles to
  0x100893 cycles)
- 3% reduction in binary size (p384_sign.elf goes from 5016 bytes to
  4872 bytes, according to riscv-unknown-elf-size)

The binary size reduction is mostly space saved by not storing the Barrett
constant; the imem size only shrinks by 16 bytes.

Alternatively, using the more generic Solinas implementation for both p
and n would result in a smaller code size but less speedup:
- 1.21x speedup for p384_ecdsa_sign_test (from 0x1c3ec6 cycles to
  0x176a35 cycles).
- 1.19x speedup for p384_ecdsa_verify_test (from 0x142f77 cycles to
  0x10e46a cycles)
- 6% reduction in binary size (p384_sign.elf goes from 5016 bytes to
  4732 bytes, an additional 156 bytes of instruction memory saved)

Signed-off-by: Jade Philipoom <jadep@google.com>