Add DSQL loader operations reference

[amaksimo]

Summary

The current DSQL skill documents the loader at a basic invocation level (~30 lines in connectivity-tools.md) but does not cover the operational knowledge that determines whether a load succeeds and at what rate. This PR adds a dedicated references/data-loading.md with the operational details an agent needs when planning or diagnosing a load.

New file: plugins/databases-on-aws/skills/dsql/references/data-loading.md

Topics covered:

• Fresh-vs-warm partition behavior — DSQL tables start on a single partition; sustained writes are required to drive splits; a fresh table absorbs roughly 3-4K rec/s from a single client. This is DSQL-side behavior, so any client-side tuning advice without it leads to misdiagnosis.
• Resume and retry mechanics — --manifest-dir, --resume-job-id, --keep-manifest. Includes the Amazon Linux 2023 /tmp tmpfs gotcha that silently makes resume impossible across an OOM/SIGKILL/reboot.
• --on-conflict do-nothing semantics — when it is safe to use and a common pitfall (duplicate-PK rows in the source).
• Schema inference caveats — when --dry-run should be used to surface the inferred schema first.
• Index count → throughput — the 1 + num_indexes write cost and when to defer index creation with CREATE INDEX ASYNC.
• Diagnostic decision tree — five symptom→cause mappings for slow or unexpected loads.

The existing loader entry in connectivity-tools.md keeps its install/quickstart content and now links out to the new reference. SKILL.md gets a one-entry index update.

Test plan

• Render the new markdown locally; verify all internal links resolve (scaling-guide.md, development-guide.md, connectivity-tools.md).
• Confirm CREATE INDEX ASYNC references match existing usage in the skill.
• Confirm no project-specific or vendor-specific content snuck in (the content was distilled from a separate cost-estimator project; only DSQL/loader knowledge that applies to any user was kept).