Fix conda packaging: Python 3.14 ABI and version solver preference

[kmarchais]

Problem

Closes #75

Two issues with the conda packages on Linux:

1. Python 3.14 free-threaded ABI (cp314t)

The simcoon conda package for linux-64 + Python 3.14 is built against the free-threaded Python variant (cp314t) instead of the standard variant (cp314). This makes it impossible to install simcoon alongside packages like matplotlib on Linux.

This is caused by a known conda-build bug where conda build on Linux incorrectly resolves python=3.14 to the free-threaded variant.

Platform	simcoon py314 ABI	Status
linux-64	python_abi 3.14.* *_cp314t	Wrong (free-threaded)
osx-arm64	python_abi 3.14.* *_cp314	Correct (standard)
win-64	python_abi 3.14.* *_cp314	Correct (standard)

2. Solver prefers simcoon 1.9.6 over 1.10.1

When Python 3.13 is installed from the defaults channel and the user runs conda install -c conda-forge -c set3mah simcoon, the solver picks 1.9.6 instead of 1.10.1.

Root cause: conda-build's automatic run_exports from libblas/liblapack in the host environment bake a tight version constraint (libblas >=3.11.0,<4.0a0) into the package metadata. This forces the solver to pull in mkl-2025.3.0 from conda-forge (instead of the older mkl-2022.1.0 from defaults), which also requires superseding _openmp_mutex. The solver with flexible priority avoids this disruption by falling back to 1.9.6.

Then 1.9.6's boost >=1.84 dependency pulls in libboost/ICU, which conflicts with matplotlib's ICU requirements.

Fix

python_abi pin (commit 0b559f1)

host:
  - python_abi {{ python }} *_cp{{ python | replace(".", "") }}

Forces the standard CPython ABI for each Python version, avoiding the free-threaded variant.

ignore_run_exports (commit c228a7e)

build:
  ignore_run_exports:
    - libblas
    - liblapack
    - libcblas

Prevents the auto-generated version constraints from being added to the package metadata. The explicit run dependencies (libblas * *mkl, etc.) already correctly pin the BLAS backend — the BLAS/LAPACK C/Fortran ABI is stable across versions, so version pinning is unnecessary.

Test plan

• Verify rendered metadata has no libblas/liblapack version constraints
• Build locally and confirm index.json dependencies are correct
• End-to-end test: conda create with Python 3.13 from defaults, install fixed simcoon -> solver picks 1.10.1 (not 1.9.6)
• End-to-end test: conda install matplotlib succeeds after installing fixed simcoon
• CI: rebuild conda packages on all platforms (linux-64, osx-arm64, win-64)
• CI: verify coinstallation with matplotlib passes

[chemiskyy]

The python_abi pin is considered to be permanent or a temporary fix? If not, we should add a comment about it for a future removal

[kmarchais]

The python_abi pin is considered to be permanent or a temporary fix? If not, we should add a comment about it for a future removal

If the linked conda-build issue is resolved in a future version, we could remove it. I added a comment next to this line to mention it, thanks!