VT_SAMPLE_SPLIT is a new production env lever with no docs/ENVIRONMENT.md row

Row: BACKEND-ROCM

#3010 adds a new production env lever, VT_SAMPLE_SPLIT, read in
src/vt/rocm/rocm_sample.hip, and documents it nowhere. The change touches one
file, so it reaches neither docs/ENVIRONMENT.md nor
scripts/env-doc-allowlist.txt.

scripts/check-env-doc.py reds on the merged result:

ERROR: production env var(s) read from src/+include/ are neither documented in
docs/ENVIRONMENT.md nor on scripts/env-doc-allowlist.txt:
  - VT_SAMPLE_SPLIT

It should be documented rather than allowlisted. The allowlist is for
kernel-internal tuning switches, and this one is neither internal nor inert: the
pull request's own A/B evidence is stated in terms of it (VT_SAMPLE_SPLIT=0 vs
1 over 20 seed/temperature combinations), it selects between two sampling
implementations on the default decode path, and 0 is the documented way to get
the pre-change kernel back for a same-binary comparison. That is a user-facing
knob by every test the reference applies.

Nothing else about the change is in question. The split-phase reduction reuses
ArgReduce, which compares true global indices rather than thread order, so
bit-identity including ties follows from the same property #2778 established, and
the 20-combination A/B plus the three-run determinism check are the right
evidence for it.

Found while gating #3010. Fixed in the same flow on that pull request's branch.

Start with docs/ENVIRONMENT.md, scripts/env-doc-allowlist.txt, and scripts/check-env-doc.py, then inspect the VT_SAMPLE_SPLIT reference in src/vt/rocm/rocm_sample.hip. Done means the production variable has a documentation row rather than an allowlist entry, and the environment documentation check passes.