Add Inf/-Inf regression tests for fire module

[brendancol]

What

Adds 28 tests covering +Inf / -Inf inputs for every public function in xrspatial.fire: dnbr, rdnbr, burn_severity_class, fireline_intensity, flame_length, rate_of_spread, and kbdi.

Why

The fire kernels skip cells only on NaN (v != v), so non-finite-but-not-NaN values flow straight through the arithmetic. That path had no test. Probing all four backends shows the behavior is consistent and well defined, for example:

• dnbr: inf - inf -> nan, inf - finite -> inf
• burn_severity_class: +inf -> class 7, -inf -> class 1
• kbdi: prev = inf clamps to 800
• rate_of_spread: slope = inf -> nan

The new tests lock this contract so a later change to the finite-value guard cannot alter behavior unnoticed.

Coverage

Each function gets a numpy contract test (exact expected values) plus four-backend parity (numpy / cupy / dask+numpy / dask+cupy). All 28 ran and passed on a CUDA host. Full test_fire.py is green (101 passed).

Test-only change. No source behavior is modified.

From the test-coverage deep sweep (Cat 2: NaN/Inf edge cases).

[brendancol]

Review: Inf/-Inf regression tests for fire

Reviewed on a CUDA host. Test-only change: 28 new tests plus a state-CSV row. No source files touched. I re-ran the file (101 passed) and checked each locked contract value against the kernels.

Blockers

None.

Suggestions

None in the added code. I hand-checked every expected value in _INF_CONTRACTS against the kernel arithmetic and all seven match:

• dnbr: inf - inf is nan, finite - inf is -inf, so [inf, -inf, -inf, nan] is right.
• rdnbr: p = inf sends the denominator to inf, so d / sqrt(inf) is 0 in column 3.
• burn_severity_class: +inf falls through to class 7, -inf lands in class 1.
• rate_of_spread: moist = inf makes eta_M nan (via inf - inf in the polynomial), so that cell is nan; wind = inf stays inf.
• kbdi: prev = inf and temp = inf both hit the 800 clamp; precip = inf zeroes the deficit, which then re-accumulates to ~43.96.

Nits

• The kbdi numpy contract pinned the literal 43.95906 with no explanation. Addressed in 63ad5d9 with a comment noting it is the post-clamp drought-factor re-accumulation at temp=30, annual_precip=1500.
• Pre-existing lint in this file, not from this PR and left alone to keep the change surgical: unused imports (functools.partial, math.exp, math.tan, ANDERSON_13, three assert_numpy_equals_* helpers), an unused func lambda at line 190, and an unused shape at line 533. Better as a separate cleanup than bolted onto a coverage PR.

What looks good

• All four backends covered with the usual decorators. The cupy and dask+cupy legs actually ran on the GPU here, not skipped.
• The parity tests reuse the same call builders as the numpy contract, so there is one definition per function.
• The header comment says why the tests exist (the kernels only guard on v != v, so non-finite values pass straight through) and what change would break them.

Checklist

• Backend parity: covered, 4 backends, GPU-validated
• Inf handling: this is the coverage being added
• Edge cases: Inf/-Inf now tested per function
• No source behavior change: confirmed
• Docstrings/README: not applicable, test-only
• flake8: added code is clean at max-line-length 100