feature: speed up unit tests

[victorapm]

This PR reduces setup, allocation, launch, and repeated numerical-derivative overhead in the slow unit tests while preserving the existing test cases and assertions.

Main changes:

• Disable per-allocation memory-log callbacks only for the testCRSMatrix target.
• Batch the Drucker–Prager and Modified Cam Clay load-step work.
• Batch fluid numerical-derivative evaluations so model setup is reused.
• Reuse inverse-capillary-pressure fixture setup across all 246 parameterized cases.
• Keep the optimized tests buildable and runnable in both CPU-only and CUDA configurations.

The “before” values are from the original test run. The “after” values are sequential wall-clock measurements from the CPU-only RelWithDebInfo build (clang++-23, ENABLE_CUDA=OFF, RAJA_ENABLE_CUDA=OFF).

Test	Before	After	Improvement
testInverseCapillaryPressure	40.11 s	0.07 s	573x faster, 99.8% less time
testModifiedCamClay	225.40 s	<0.01 s	>22,540x faster, >99.99% less time
testMultiFluidBlackOil	15.82 s	0.07 s	226x faster, 99.6% less time
testMultiFluidCO2Brine	79.88 s	2.84 s	28.1x faster, 96.4% less time
testMultiFluidThreePhaseCompositionalMultiphase	44.89 s	0.11 s	408x faster, 99.8% less time
testMultiFluidTwoPhaseCompositionalMultiphase	115.92 s	0.22 s	527x faster, 99.8% less time

Additional CPU-only timings: testCRSMatrix 0.02 s, testDruckerPrager 0.01 s, and testMultiFluidDeadOil 0.07 s.

• CPU-only build: 534/534 targeted tests passed; the slowest targeted test took 2.84 s.
• CUDA build: 573/573 targeted tests passed, including the host/device variants.
• Tests were run sequentially because several fluid tests write shared CSV filenames.

Depends on GEOS-DEV/LvArray#355