Fix bit_length=64 undefined behavior, make it configurable, and lower --samples/--max_nb defaults

[hfwen0502]

1. bit_length = 64 is undefined behavior

bit_length sets how many bits SBD packs into each size_t of its
std::vector<size_t> bitstring representation. The Python wrapper hardcoded 64
in four places, but bitadvance() in framework/bit_manipulation.h computes

size_t d = (((size_t) 1) << bit_length) - 1;

Shifting a 64-bit size_t by 64 is undefined behavior. In practice the shift
count is masked to 0, so the mask collapses to d == 0 (checked on aarch64 and
x86-64). bitadvance() is reached from mpi_redistribution() and
mpi_sort_bitarray() — the multi-rank determinant distribution paths.

RIKEN documents the default as 20 (see --bit_length in
apps/chemistry_tpb_selected_basis_diagonalization/README.md), which is also
what run_sbd_diag.py already defaults to. Only run_sqd_sbd.py passed 64.

This adds SBD_DEFAULT_BIT_LENGTH = 20 and a --bit_length option. The value
is not range-checked: callers can still pass 64 explicitly if they want the old
packing.

2. Latent correctness bug: the value only reached half the pipeline

bit_length decides both how Python packs determinants and how the C++ engine
interprets them, so the two must agree. Previously a caller-supplied
bit_length reached the engine via TPB_SBD.bit_length but not
from_string()/makestring(), which stayed at 64 — so --bit_length 48 would
have silently corrupted every determinant.

_ci_strings_to_sbd_dets() and _sbd_dets_to_ci_strings() now take
bit_length, and _solve_sci_core() builds the config before packing so both
sides read one value (sbd_data.bit_length, the same object handed to
tpb_diag).

3. Defaults and example

• --max_nb 50 → 10, matching upstream chemistry/tpb/sbdiag.h.
  davidson_thrust.h allocates 2 * num_block device vectors of length dim,
  so GPU workspace scales as dim * (2 * max_nb + ~4) * 8 bytes — roughly a 4x
  reduction.
• --samples 10000 → 3000.
• run_sqd_sbd.ipynb sbd_config updated from 64 to 20 (its max_nb was
  already 10). The notebook is executed by the nbmake tox env in CI.

Validation

Threading actually verified (GB200, _create_sbd_config + converters called
directly). Note that comparing energies across bit_length values cannot prove
this on its own: a silently-ignored flag also yields identical energies. These
two checks distinguish the cases.

Config reaches the object passed to tpb_diag:

no override                    -> sbd_data.bit_length = 20
sbd_config={'bit_length': 48}  -> sbd_data.bit_length = 48
sbd_config={'bit_length': 63}  -> sbd_data.bit_length = 63

Packing honours it — norb=24, so word count must be ceil(24/bit_length):

bit_length=20 -> 2 size_t words   round-trip OK
bit_length=48 -> 1 size_t word    round-trip OK
bit_length=63 -> 1 size_t word    round-trip OK

The changing word count is the observable that rules out the flag being dropped,
and each det round-trips back to its original CI string.

End-to-end energy (NVHPC 25.5, cc100, 4 GPUs, CUDA-aware MPICH 5.0.1):
H2O 1e-5 subspace, dim 303,601 → -76.2437350421 at both np=1 and 2x2,
matching the pre-change value and run_sbd_diag.py's independent file-based
path (-76.2437349413). bit_length 20/48/63 all agree to the last digit.
norb=24 at the new default spans two size_t words, so this exercises the
multi-word packing path (previously always single-word at 64).

Not verified

• The new --max_nb 10 default is upstream's value and reduces memory as
  described, but has not been checked for convergence on a large production
  system; a smaller Krylov space may need more restarts.
• The H2O energies above were produced at max_nb = 50; the new default was not
  separately re-validated.