Relax over-tight dependency floors and single-source the ruff pin

[Panadestein]

Problem

Over-tight runtime pins

numpy = ">=2.4.4,<3"
quimb = ">=1.14.0,<2"

numpy>=2.4.4 demands a NumPy released only weeks ago. A library should
declare the oldest version it actually works with, not the newest one that
happened to be installed when the constraint was written. As it stands,
src_method cannot be co-installed with anything that has not yet moved to the
very latest NumPy, which is a real adoption barrier for a scientific package —
users typically have a large pinned environment they cannot churn.

Nothing in the source uses a NumPy 2.4-only feature; the code is compatible far
further back.

Tool version drift

The ruff version is pinned in two places that disagree:

• pyproject.toml dev group: ruff==0.15.12
• .pre-commit-config.yaml: rev: v0.12.11

So pre-commit and uv run ruff can produce different results on the same
file — a hook passes locally and CI fails, or vice versa. There is an
update-ruff.yml workflow that appears to maintain one of them but not the
other.

Proposed fix

• Lower the NumPy floor to the oldest version the test suite actually passes
  on. Determine it empirically rather than guessing: run the suite against
  candidate floors (2.0, 2.1, 2.2) in a matrix and set the constraint to
  the lowest that is green. Keep the <3 major-version cap.
• Do the same for quimb, whose floor was likewise set to a very recent
  release.
• Make ruff single-sourced: either have update-ruff.yml bump both locations
  atomically, or drop the version from one of them so there is a single source
  of truth. Add a CI check that fails if the two ever diverge.
• Add the oldest supported dependency set to the CI matrix so the floor is
  actually tested rather than merely asserted.

Acceptance criteria

• NumPy and quimb floors reflect tested compatibility, not the latest release.
• CI exercises the declared minimum dependency versions.
• ruff is pinned to the same version in pyproject.toml and .pre-commit-config.yaml, enforced automatically.