Restore accuracy truncation bound in 2.1.2

[shinaoka]

Summary

• restore the sparse-ir 1.x meaning of FiniteTempBasis.accuracy: the relative first excluded singular value
• retain the last-included fallback only when the full SVE has no excluded value
• cover epsilon truncation, max_size truncation, and the no-excluded-value case
• allow the upcoming pylibsparseir 0.9.x release and advertise Python 3.14 support
• fix the README filename used by package metadata and the source manifest
• bump sparse-ir to 2.1.2

Why

In 2.x, accuracy returned the last included singular value. That reversed the relationship expected by callers using requested_tolerance > basis.accuracy. The first excluded singular value is the truncation error bound and matches sparse-ir 1.x behavior.

Verification

• Python 3.11: 97 passed, 8 skipped
• Python 3.14 with locally built pylibsparseir 0.9.0 wheel: 97 passed, 8 skipped
• source distribution and universal wheel build successfully
• pylibsparseir dependency specifications are consistent between PyPI and conda metadata

Release ordering

This PR accepts pylibsparseir 0.9.x, but Python 3.14 installation requires the CPython 3.14 wheel prepared in SpM-lab/sparse-ir-rs#251. Publish pylibsparseir 0.9.0 first, then sparse-ir 2.1.2.