Add pygraspa backend for ML-potential GCMC

[tdpham2]

Summary

• Adds a new `matkit.pygraspa` backend that produces simulation directories ready to run with pygRASPA — gRASPA Monte Carlo driven by ML interatomic potentials (FAIRChem eSEN/UMA/AllScAIP, MACE-Polar, Allegro) instead of classical force fields.
• API mirrors `matkit.graspa`: `setup_simulation`, `setup_batch`, `get_output_data`, plus a `compute_ecomp` helper for the per-adsorbate reference energy pygRASPA needs in its Metropolis criterion.
• Setup-only integration — matkit writes a small `run.py` launcher into each sim dir and does not import pygRASPA / torch / CUDA at install time. No GPU dependency added to matkit itself.
• Includes single + batch setup, a single/mixture isotherm orchestrator in `scripts/pygraspa/`, `matkit pygraspa` CLI commands, and a no-GPU pytest suite.

What's in here

Area	Files
Core	`src/matkit/pygraspa/{init.py,pygraspa.py}`
Templates	`src/matkit/pygraspa/files/{template,template_mixture,template_mixture_isotherm}/` (duplicated from `matkit.graspa` — same gRASPA input format)
CLI	`matkit pygraspa setup|batch-setup|compute-ecomp|analyze` added to `src/matkit/cli.py`
Isotherm	`scripts/pygraspa/{setup_isotherms.py,isotherm_config.yaml}` (single + mixture modes, parallel to `scripts/graspa/`)
Packaging	New `pygraspa` extra + 3 `package-data` entries in `pyproject.toml`
Tests	`tests/test_pygraspa.py` (17 tests, no GPU/torch required)

Modes supported

• `mode='run'` → renders a `run.py` calling `pygRASPA.main.run` (plain ML MC).
• `mode='run-auto'` (default) → renders a `run.py` calling `pygRASPA.auto.main.run_auto` (JSON checkpoint + STOP_FLAG for preemptible cluster jobs).
• `run-classic` is intentionally omitted — `matkit.graspa` already covers that use case.

Test plan

• `pytest tests/test_pygraspa.py` — 17 new tests pass
• `pytest` (full suite, ignoring GPU-only test_uma) — 164 passed, 1 skipped, no regressions
• `ruff check` clean on all new/modified files
• CLI smoke: `matkit pygraspa setup …` produces a sim dir with rendered `simulation.input` (no placeholders left) and a syntactically valid `run.py` (`python -m py_compile run.py`)
• CLI smoke: `matkit pygraspa batch-setup …` produces correct `simulations.jsonl` and `cif_mapping.json` for 2 CIFs × 2 T × 2 P = 8 sims
• Isotherm orchestrator dry-run, single mode: 4 pressures × 1 T × 2 adsorbates = 8 sims in adsorbate-split subdirs
• Isotherm orchestrator dry-run, mixture mode: `MolFraction` overrides propagate into `Component` blocks, `E_comps` matched in order in `run.py`
• Run an actual `pygraspa run-auto` job on a GPU box and confirm `matkit pygraspa analyze` parses the resulting JSON-lines log. The parser is built against the schema described in pygRASPA's README and uses tolerant field-name matching, but should be validated against a real `output.log` before relying on it in production.