HPC fixes: band rebind bug, std::min optimization, branchless early abandon

WDTW banded rolling buffer + CI wheel fix:
- WDTW banded: full ScratchMatrix (128MB) → rolling column buffer (32KB)
  for n=4000. Matches dtwBanded pattern. Added convenience overloads
  taking g parameter instead of precomputed weights.
- CI wheels: fix CIBW_CONFIG_SETTINGS semicolon issue — separate each
  cmake arg on its own line for scikit-build-core
- nanobind CMake: try pip-installed nanobind first via cmake_dir,
  fall back to CPM fetch for standalone builds

ADTW banded kept as ScratchMatrix (agent's rolling buffer rewrite had
correctness bugs — will revisit in a future PR with proper TDD).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: ElektrikAkar

.github/workflows/python-wheels.yml

@@ -27,9 +27,12 @@ jobs:
           CIBW_SKIP: "*-musllinux_* *-win32 *-manylinux_i686"
           CIBW_ARCHS_MACOS: "x86_64 arm64"
 
-          # Disable commercial/heavy solvers for wheel builds
+          # Disable commercial/heavy solvers for wheel builds.
+          # Each cmake arg must be a separate config-setting line for scikit-build-core.
           CIBW_CONFIG_SETTINGS: >-
-            cmake.args="-DDTWC_BUILD_PYTHON=ON;-DDTWC_ENABLE_GUROBI=OFF;-DDTWC_ENABLE_HIGHS=OFF"
+            cmake.args=-DDTWC_BUILD_PYTHON=ON
+            cmake.args=-DDTWC_ENABLE_GUROBI=OFF
+            cmake.args=-DDTWC_ENABLE_HIGHS=OFF
 
           # macOS needs libomp for OpenMP
           CIBW_BEFORE_BUILD_MACOS: "brew install libomp || true"