Fix NCCL warm-up allreduce count mismatch

[rwakizaka]

Problem

When the number of determinants is not evenly divisible by bdet_comm_size, the
process reproducibly hangs inside the NCCL warm-up ncclAllReduce call in
sbd/chemistry/tpb/sbdiag.h, before "NCCL communicators have been created." is
printed.

Root cause

After init_nccl_comm, a warm-up nccl_allreduce was called with a buffer sized
W.size():

if (mpi_size_b > 1) {
    init_nccl_comm(&b_nccl_comm, b_comm);
    thrust::device_vector<double> A(W.size(), 0.0);  // ← here
    nccl_allreduce(A, ncclSum, b_nccl_comm);
}

W is the local wavefunction chunk distributed across the β-det communicator. When
the total number of determinants is not evenly divisible by bdet_comm_size, each
rank holds a different-sized slice of W. NCCL's ncclAllReduce requires all
participating ranks to pass the same element count; mismatched counts cause a
protocol deadlock.

Fix

Replace the variable-size warm-up buffers with a constant-size buffer of 1 element
for all four NCCL communicators (h, b, t, a):

if (mpi_size_b > 1) {
    init_nccl_comm(&b_nccl_comm, b_comm);
    thrust::device_vector<double> A(1, 0.0);  // fixed: all ranks agree on count=1
    nccl_allreduce(A, ncclSum, b_nccl_comm);
}

The warm-up exists solely to trigger NCCL channel setup before the main computation;
the data transferred is irrelevant, so a single-element buffer is sufficient.

Files changed

• include/sbd/chemistry/tpb/sbdiag.h — warm-up allreduce buffer size changed from
  W.size() to 1 for all four NCCL communicators