Optimize fermions

[mtfishman]

Related to #1712, these are all of the optimizations that AI came up with. This version removes all extra allocations in the fermion code.

Closes #1709, closes #1712.

[mtfishman]

The time overhead is now down to about 5-10% for a Hubbard model DMRG calculation on a length 10 chain with 6 sweeps ramping up to a bond dimension of 200, and the allocation overhead is minimal. The overhead decreases with larger bond dimensions (or more specifically, larger block dimensions) and system sizes.

Some example runs for that case:

julia> @time mps_dmrg_bench(; use_auto_fermion = false)
mps.maxlinkdim(H_mps) = 6
After sweep 1 energy=-11.969543558445253  maxlinkdim=10 maxerr=1.49E-03 time=0.034
After sweep 2 energy=-12.116369527016067  maxlinkdim=20 maxerr=1.37E-04 time=0.128
After sweep 3 energy=-12.118011427386474  maxlinkdim=40 maxerr=1.14E-05 time=0.213
After sweep 4 energy=-12.118221850986792  maxlinkdim=80 maxerr=5.02E-08 time=0.260
After sweep 5 energy=-12.118222536641227  maxlinkdim=160 maxerr=2.75E-10 time=0.323
After sweep 6 energy=-12.118222538380001  maxlinkdim=200 maxerr=9.94E-11 time=0.359
DMRG energy = -12.118222538380001
  1.325985 seconds (14.83 M allocations: 2.267 GiB, 12.80% gc time)

julia> @time mps_dmrg_bench(; use_auto_fermion = true)
mps.maxlinkdim(H_mps) = 6
After sweep 1 energy=-11.96954355844525  maxlinkdim=10 maxerr=1.49E-03 time=0.031
After sweep 2 energy=-12.116369527016055  maxlinkdim=20 maxerr=1.37E-04 time=0.135
After sweep 3 energy=-12.118011427386506  maxlinkdim=40 maxerr=1.14E-05 time=0.231
After sweep 4 energy=-12.118221850986796  maxlinkdim=80 maxerr=5.02E-08 time=0.275
After sweep 5 energy=-12.11822253664123  maxlinkdim=160 maxerr=2.75E-10 time=0.341
After sweep 6 energy=-12.118222538379994  maxlinkdim=200 maxerr=9.94E-11 time=0.373
DMRG energy = -12.118222538379994
  1.392076 seconds (14.83 M allocations: 2.298 GiB, 10.65% gc time)