Distribute external_file particle reading across all MPI ranks

[Noerr]

Summary

AddPlasmaFromFile() previously loaded the entire openPMD particle file on the IO rank, causing GPU OOM for large files. This PR distributes the file reading across all MPI ranks using loadChunk(offset, extent) with sub-chunking to bound peak memory.

• All ranks open the file collectively via openPMD::Series(path, READ_ONLY, Communicator())
• Particle index range divided evenly: rank_chunk = ceil(npart_total / nranks)
• Sub-chunk loop with max_sub_chunk = 2^22 (4M particles) bounds peak memory at ~960 MB/rank
• Each iteration: loadChunk(offset, extent) → insideBounds() filter → collective AddNParticles()
• series.flush() called outside the sub_count > 0 guard so all ranks participate (flush may be collective with ADIOS2/HDF5 MPI backends)
• Added WARPX_ALWAYS_ASSERT_WITH_MESSAGE to catch particle overcount bugs

Follows the same distributed I/O pattern established by PR #6221 (distributed density reading).

Changes

File	Description
PlasmaInjector.H	Replace std::any m_openpmd_input_series with std::string m_injection_file_path
PlasmaInjector.cpp	Store file path instead of Series handle; Series closes after metadata read
AddParticles.cpp	Distributed sub-chunked reading in AddPlasmaFromFile()

Motivation

• Reproduction: Frontier (OLCF), 64 MI250X GCDs (62 GiB HBM each), 117 GB openPMD/HDF5 file (~2.4B particles) → hipMalloc returned 2: out of memory on rank 0
• Root cause: IO rank must hold ~133 GB of particle data in memory before Redistribute(), exceeding GPU memory
• TODO comment: //TODO: Make changes for read/write in multiple MPI ranks has been in the code since PR Load Particles: external_file MPI Support #956 (May 2020)

Validated on Frontier (AMReX TinyProfiler)

Test case: 4 nodes × 8 GCDs = 32 MPI ranks, ~740M particles.

Metric	Serial (IO rank only)	Distributed (this PR)
GPU AddParticles() MaxMem max	31 GiB	701 MiB (45× reduction)
Pinned AddParticles() MaxMem max	63 GiB	495 MiB (130× reduction)
GPU Arena used (min...max)	878 MiB ... 46,065 MiB	1,101 MiB ... 3,188 MiB
Pinned Arena used (min...max)	381 MiB ... 97,277 MiB	903 MiB ... 1,297 MiB
Redistribute_partition Nalloc	194,517	7,835

The serial run had rank 0 at 96% of GPU arena capacity. The distributed approach keeps all ranks under 3.2 GiB.

Test plan

• Existing test test_3d_focusing_gaussian_beam_from_openpmd_picmi should pass (exercises FromFileDistribution with 2 MPI ranks)
• Verified on Frontier: 32 MPI ranks, ~2.4B particle file, no OOM, simulation ran to completion where it previously failed on particle load.
• Verified on Frontier: 32 MPI ranks, ~740M particle file. Simulation produces same answer over ~3000 time steps for the legacy serial load and this distributed particle load.
• Confirm CI passes (eg. builds 1D, 2D, 3D, RZ)

Resolves #3185