Skip tautological compiled programs in RuleCache

[snuderl]

Summary

• Mirrors protovalidate-go's ReduceResiduals optimization in RuleCache: after compiling a rule's CEL program, we eagerly evaluate it with only the rules / rule variables bound.
• If the program returns true or "" (the success values), it is tautological for this rule configuration and is dropped before being added to the per-field program list, so it never runs during validation.
• Programs that depend on the unbound this (or now) variables raise CelEvaluationException during the probe; we swallow that and keep the program as before.

This is a follow-up to #454, which performed a similar tautology check at the MessageEvaluator level.

Test plan

• ./gradlew test — full unit + conformance suite passes locally.

[snuderl]

Benchmark results

Added a NumericRangeMessage benchmark targeting this optimization: 20 float fields with only gt set, so 4/5 of the float.gt rule expressions short-circuit to '' using only rules and become tautological.

A/B run on the same machine, JDK 21, JMH defaults (3×2s warmup, 5×2s measurement, 2 forks, AverageTime mode). Baseline is bufbuild/main + benchmark commit only; "after" adds the RuleCache change.

benchmark             metric  before          after           delta
validateNumericRange  time    28429.09 ns/op  17028.83 ns/op  -40.1%
validateNumericRange  alloc   97920.01 B/op   58560 B/op      -40.2%
validateManyUnruled   time      469.29 ns/op    453.31 ns/op   -3.4%
validateManyUnruled   alloc     1992    B/op   2044    B/op    +2.6%
validateRegexPattern  time     1504.76 ns/op   1545.29 ns/op   +2.7%
validateRegexPattern  alloc     3944    B/op   4072    B/op    +3.2%
validateRepeatedRule  time     9878.88 ns/op  10196.05 ns/op   +3.2%
validateRepeatedRule  alloc    40289.13 B/op   41600    B/op   +3.3%
validateSimple        time     1519.79 ns/op   1544.66 ns/op   +1.6%
validateSimple        alloc     3520    B/op   3640    B/op    +3.4%

The targeted benchmark drops ~40% in both time and per-op allocation (the four eliminated CEL programs are no longer evaluated and don't allocate intermediate result objects). Other benchmarks move within the typical run-to-run noise band for the default iteration budget — none of them have rule expressions that residualize to a constant, so no real change is expected there.

[jonbodner-buf]

Are there other rules that expand in the same way as gt/lt?