Add Execution Models docs and reorganize parallelism/GIL

[mthrok]

Restructure the pipeline-parallelism documentation into a fundamentals page, a patterns page, and a GIL page, and align the surrounding guides.

• Pipeline Parallelism (getting_started/parallelism.rst) is refocused as the fundamentals: the executor mechanism, dedicated thread pools (e.g. GPU transfer), the cost of crossing a process boundary, the .to() region API, and run_pipeline_in_subprocess. The obsolete "Multi-processing (combined)" section is removed; its point — a per-stage ProcessPoolExecutor pays a data round trip at every stage boundary — is folded into a new "cost of crossing a process boundary" subsection, which also covers cheap large-payload transfer (prefer torch.Tensor / NumPy arrays, which move through shared memory; wrap a raw byte string in a 1-D uint8 tensor; SharedMemorySegmentPool for the last bit of IPC performance). A note explains that a region stage's concurrency applies within each worker, so effective concurrency across the pool is concurrency × max_workers.
• Execution Models (getting_started/execution_models.rst, new) documents the three patterns built on those primitives — MT, MTP, and MP — each with an ASCII diagram, a minimal PipelineBuilder snippet (using a continuous source), and guidance on when to use it. It introduces the noisy-neighbour effect (the loader's CPU work delaying the GPU kernel launches issued from the training process) and recommends MTP, and an intermediate-process MP, for GPU training.
• Working Around the GIL (getting_started/gil.rst) replaces faq.rst (which had become GIL-only): which operations release the GIL, a pandas-vs-polars example (with a figure), what to do when a function does not release it, and which functions hold it.

Supporting changes:

• The PipelineBuilder.to, SharedMemorySegmentPool, and SharedMemoryRingBuffer docstrings gain the region-concurrency note and links to the benchmark_arena_transport example / shared-memory-arena case study.
• pytorch.rst links the new pages and closes with a short "isolate the loader (MTP)" pointer for production, while keeping its focus on decomposing a Dataset into stage functions.
• The obsolete with pipeline.auto_stop() wrapper is removed from doc examples (the section that documents auto_stop itself is kept).

GIL guidance is de-duplicated so gil.rst is the single canonical source and the other pages link to it. Toctrees and cross-references are updated; the existing pipeline-parallelism and pipeline-parallelism-custom-mt labels are preserved so inbound links keep working.

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D112516147. (Because this pull request was imported automatically, there will not be any future comments.)