numerical (?) issues with multithreaded runs

[nmayhall-vt]

Running the same calculation with either -t1 or -tauto is giving small differences in even the first iteration when starting from single vectors from multiple roots

 ===================================================================
     Selected CI Iteration:    1 epsilon:   0.00100000
 ===================================================================
 Build Hamiltonian matrix with dimension:    25
  0.001247 seconds (15.92 k allocations: 326.172 KiB)
  Root       Energy           S2
     1 -18.32003824   0.00000000
     2 -17.65879673   1.99998754
     3 -17.65230384   1.99988141
     4 -17.65098494   0.00011089

 ===================================================================
     Selected CI Iteration:    1 epsilon:   0.00100000
 ===================================================================
 Build Hamiltonian matrix with dimension:    25
  0.001236 seconds (15.92 k allocations: 326.172 KiB)
  Root       Energy           S2
     1 -18.32003824   0.00000000
     2 -17.65881088   2.00000000
     3 -17.65231530   2.00000000
     4 -17.65101589   0.00000000

input file attached
test_matvec.txt

[nmayhall-vt]

This also exists for BST. The first inconsistency occurs during the matvec

It's probably an issue of the order that the sigma Tuckers are added. Maybe collect all and do 1 SVD on the super-wide matrices to find the span.