NEWS/test cleanup: correct the dontrun count, add a bf16 source check

[TroyHernandez]

Review follow-ups, all documentation or test scope. No code changes, so no
version bump.

Doc fixes

• NEWS.md said nine \dontrun{} examples remain while cran-comments.md
  said ten. Ten is right: the 0.2.0 review pass left nine, and
  resident_load() added one afterwards.
• Dropped a doubled blank line left by consolidating the dev sections under
  0.2.2, and a trailing blank line at the end of test_recommend.R.
  git diff --check is clean.

bf16 coverage

The bf16 route had only structural coverage — that .flux_model_dir() routes
bf16 somewhere other than "<prefix>bf16", and that an unknown model errors.
Nothing checked what actually carries risk: whether .flux_source_dir() lands
on a directory the loader can open.

test_bf16_source.R asserts the cache layout (config.json plus a shard),
that the router resolves to exactly that directory, that every tensor in the
shard really is BF16 — read from the safetensors JSON header, so no tensors
are materialized — and that flux_open_checkpoint() opens it and reports no
format, which is what selects the "full" bf16 path in
flux_load_transformer().

Deliberately not a forward pass. bf16 names the storage dtype and reading
it is CPU work, but computing in bf16 is GPU-only here (txt2img_flux2()
upcasts to float32 when device == "cpu"), and materializing the 7.8 GB
transformer is not a test-suite job. at_home() gates the file so check
machines skip it.

Verified GPU-untouched: 11 assertions in 1.6 s with a 0 MiB delta in
nvidia-smi. Full suite with CUDA hidden: 1081 pass.