refactor(docker): drop redundant --force-reinstall, --no-deps is enough

Empirical test in a fresh devel_latest container after the prior fix
(e2b82b7) showed `pip install --no-cache-dir --no-deps` with the three
cutlass-dsl subpackages named explicitly produces the correct PyPI
4.4.1 wheel content (3241-byte cute/arch/__init__.py with the
top-level ProxyKind / SharedSpace re-export) without --force-reinstall.

The reason --no-cache-dir alone fails (still gets the cached 3232-byte
wheel) but --no-deps + --no-cache-dir + explicit subpackages succeeds
is that `[cu13]` extra routes pip through a dep-resolution code path
that doesn't fully honor --no-cache-dir, whereas explicit-args with
--no-deps takes a simpler install path where --no-cache-dir does take
effect. --force-reinstall was belt-and-suspenders we no longer need.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Signed-off-by: Runchu Zhao <zhaorunchu@gmail.com>

Author: z52527

docker/Dockerfile

@@ -39,11 +39,12 @@ RUN if [ "${TRITONSERVER_BUILD}" = "1" ]; then \
 # Megatron-LM core_v0.13.1: contains d9608004f which gates
 # ChainedOptimizer.count_zeros_fp32 on log_num_zeros_in_grad (~4 ms/step
 # saving on HSTU bf16); 0.12.x always pays the cost.
-# nvidia-cutlass-dsl: triple flag below avoids a poisoned base-image
-# pip cache wheel; keep the re-pins on the bundled install too.
+# nvidia-cutlass-dsl: --no-deps + --no-cache-dir + 3 explicit subpackages
+# avoids a poisoned base-image pip cache wheel; keep the re-pins on the
+# bundled install too.
 RUN pip uninstall -y nvidia-cutlass-dsl nvidia-cutlass-dsl-libs-base nvidia-cutlass-dsl-libs-cu13 || true && \
     rm -rf /usr/local/lib/python3.12/dist-packages/nvidia_cutlass_dsl* && \
-    pip install --no-cache-dir --no-deps --force-reinstall \
+    pip install --no-cache-dir --no-deps \
         'nvidia-cutlass-dsl==4.4.1' \
         'nvidia-cutlass-dsl-libs-base==4.4.1' \
         'nvidia-cutlass-dsl-libs-cu13==4.4.1' && \