Add pull request templates with an exhaustive new-model checklist - #1169

Closes #913.

Motivation

#913 asks for "an exhaustive todo list for adding models ... directly included in a PR template". The Adding a model to Braindecode guide in CONTRIBUTING.md covers the implementation conventions, but a contributor currently discovers the remaining requirements only through CI failures or by reverse-engineering recent model PRs:

• several conventions are enforced by test/unit_tests/models/test_integration.py without being listed anywhere upfront (self.final_layer naming, activation exposed as a class-default parameter, drop-probability parameters, summary.csv completeness);
• the registration and documentation steps (registry in models/util.py, export in models/__init__.py, summary.csv, docs/api.rst, architecture figure) are only visible in past model PRs such as Adding ZUNA to Braindecode #1020;
• benchmark evidence supporting the reported numbers has become an expectation in model reviews (Make more official way to validate models #1156), but nothing asks for it up front.

I hit exactly these gaps while preparing the MSCFormer addition for #721 — the conventions first surfaced as local test failures — which motivated this PR.

Changes

• .github/PULL_REQUEST_TEMPLATE/add_new_model.md: template for model additions with the exhaustive checklist (model information, implementation conventions annotated with the enforcing test names, registration and documentation steps, validation and benchmark expectations).
• .github/PULL_REQUEST_TEMPLATE/default.md: minimal general template (summary / changes / testing / notes). Two templates in PULL_REQUEST_TEMPLATE/ make GitHub show a template chooser; with a single file, GitHub would silently apply the model checklist to every PR, which is not intended.
• CONTRIBUTING.md: one-line pointer from the "Add a model" section to the template.
• docs/whats_new.rst: entry for the changelog check.

Design notes

• The checklist deliberately mirrors the guide rather than replacing it: it points contributors to CONTRIBUTING.md for the how and serves as the what to tick.
• The benchmark section is phrased as "reproduce the paper's results on at least one public dataset, or explain the gap" — a form-level ask that leaves the broader policy discussion (Make more official way to validate models #1156) to maintainers.
• No new dependencies; markdown/rst-only changes.

Testing

• No code changes. Both templates have valid YAML frontmatter (name / description / title), and the chooser behavior was verified to require two or more templates in the directory.
• codespell and sphinx-lint pass on the touched files (docstrfmt excludes whats_new.rst in pre-commit by design; verified the exclusion before relying on it).
• Changelog check satisfied via the whats_new.rst entry.