feat: support SQL IS [NOT] DISTINCT FROM as null-safe comparison

[jonasdedden]

Support SQL IS [NOT] DISTINCT FROM in Lance filters, which previously failed with "not supported SQL".

The planner now maps both spellings to DataFusion's IsDistinctFrom / IsNotDistinctFrom, as DataFusion's own SQL frontend does. Literal coercion, the signed-zero rewrite (x IS NOT DISTINCT FROM 0.0 matches both zero encodings) and null-safe evaluation already handle these operators. Also fixes the signed_zero.rs doc table, which described an outdated rewrite shape.

Column-vs-column signed zeros and NaN still follow Arrow's total order (#9316, fixed by #9323; #9315).

[jonasdedden]

@lance-gatekeeper: Agreed. The risk isn't specific to this spelling: = and != between float columns already follow Arrow's total order. This PR only adds the SQL syntax and doesn't change comparison semantics. The PR description now points to #9316 (signed-zero column comparisons, fixed by #9323) and #9315 (NaN ordering). I also removed the test asserts that treated the current -0.0/NaN results as correct, so those fixes won't break this PR's tests when they land.

[lance-gatekeeper]

⚠️ Gate recommendation: approve with a non-blocking risk.

The author clarification confirms that this PR only exposes the existing comparison semantics and now avoids tests that pin the signed-zero or NaN total-order behavior. The direct parser lowering still matches DataFusion and preserves null-safe evaluation plus the existing signed-zero literal rewrite, so no change is requested here. The accepted residual risk is that float column-to-column and NaN distinctness can still follow Arrow total order; #9323 is the active signed-zero repair and #9315 tracks NaN ordering.