[Docs] Document SLURM array sharding, completion tracking, and retries

[lbliii]

Context

Child of #2118. PR #2059 added first-class SLURM job-array support, deterministic source-task sharding, durable shard completion manifests, FailedTask-aware retries, and reference scripts under tutorials/slurm/. The operational guide currently exists only in the repository README and is not discoverable in the Fern documentation.

Documentation scope

Add a Fern guide under Admin > Deployment > SLURM that covers:

• when to use independent SLURM array jobs instead of one multi-node Ray cluster
• JSONL and Parquet setup with submit_array.sh and array_pipeline.py
• automatic SLURM_ARRAY_TASK_ID / SLURM_ARRAY_TASK_COUNT discovery
• NEMO_CURATOR_SLURM_ARRAY_ENABLED, shard-index, total-shards, and minimum-index overrides
• deterministic task-ID hashing and source-task assignment requirements
• single-node RayClient versus multi-node SlurmRayClient within each array task
• shared checkpoint_path, run.json, completion manifests, and retry discovery
• pipeline exceptions versus downstream FailedTask results
• why source stages cannot emit FailedTask
• array-size limits, windowed submissions, SHARD_INDEX_OFFSET, and empty shards
• attempt-scoped failed-task directories and checkpoint reuse/reset rules
• retry_array.py, including maximum physical array size handling

Link the existing tutorial scripts instead of duplicating them wholesale. Do not document nemo-curator-slurm-cli; it was explicitly not implemented in #2059.

Related updates

• Add the page to fern/versions/main.yml under the existing SLURM deployment section.
• Cross-link Multi-Node Ray on SLURM, Resumable Processing, and the admin landing page.
• Add a concise Add support for Slurm arrays #2059 feature entry to the 26.06 release notes in [codex] publish 26.06 release notes and migration checklist #2143.
• Keep generic source-level checkpoint semantics in [codex] docs: document native pipeline resumability #2134 and route array-specific operations to this guide.

Acceptance criteria

• A runnable JSONL array example reaches successful completion tracking
• Parquet configuration differences are explicit
• Environment variables and validation constraints match current source
• Deterministic sharding and non-deterministic r task-ID rejection are explained
• Completion, failure, and retry states are tabulated
• Oversized logical arrays and offset submissions are documented
• Checkpoint directory layout and attempt-scoped failure manifests are shown
• New page is in navigation and linked from related infrastructure pages
• Add support for Slurm arrays #2059 appears in 26.06 release notes
• Fern validation and link checks pass

Related implementation: #2059
Parent tracker: #2118