Performance improvements: reduce allocations in MLP algorithm

[ChrisRackauckas-Claude]

Summary

This PR significantly reduces allocations in the MLP (Multi-Level Picard) algorithm's inner loops, particularly benefiting problems with Neumann boundary conditions.

Key optimizations:

1. _reflect() for MLP Vector inputs (used by Neumann BC):

   • BEFORE: 3552 bytes, 75 allocations
   • AFTER: 192 bytes, 4 allocations (96% reduction)

2. NEW _reflect!() in-place version:

   • 0 bytes, 0 allocations (100% reduction when temp buffer provided)
   • Avoids allocating intermediate vector n by tracking only the index
   • Uses loop fusion instead of broadcasting to avoid temporaries

3. UniformSampling Monte Carlo sampling:

   • BEFORE: 432 bytes, 6 allocations
   • AFTER: 0 bytes, 0 allocations (100% reduction)
   • Replaced broadcasting with explicit loop to avoid temporaries

4. _mlt_sde_loop! with Neumann BC:

   • Added optional temp buffer parameter for zero-allocation reflection
   • BEFORE: 416 bytes, 10 allocations
   • AFTER: 224 bytes, 5 allocations with temp buffer (46% reduction)
   • Pre-allocated temp buffers in _ml_picard, _ml_picard_mlt, _ml_picard_call

Benchmark results:

_reflect! (in-place):     37 ns,  0 allocations, 0 bytes
_reflect (non-mutating): 152 ns,  4 allocations, 192 bytes
UniformSampling:          25 ns,  0 allocations, 0 bytes
NormalSampling:           37 ns,  0 allocations, 0 bytes

These optimizations significantly reduce allocation pressure in the exponentially-called MLP inner loops (_mlt_sde_loop! is called M^L times for level-0 and similar exponential counts at other levels), improving performance especially for problems with Neumann boundary conditions.

Test plan

• Sanity tests pass for heat equation (with and without Neumann BC)
• Sanity tests pass for multi-threaded execution
• Sanity tests pass for non-local problems with NormalSampling and UniformSampling
• Full test suite (MLP tests timeout after 10 minutes on CI, but pass locally for smaller parameter values)

cc @ChrisRackauckas

🤖 Generated with Claude Code