fix: align parallel context sizing with slots

[inureyes]

Summary

Implements issue #57 by aligning server parallel context allocation with llama.cpp server semantics:

• Treat explicit --ctx-size C as a total context budget shared by active request slots.
• Resolve effective per-slot context as floor(C / N) for --parallel N.
• Use an explicit --max-batch-size M as the divisor because it is the actual concurrent decode-sequence limit.
• Keep --no-batch on single-slot semantics so it receives the full explicit context budget.
• Reject startup when the effective per-slot context is below 512 tokens.
• Derive the plain KV live-window cap from the same effective per-slot context, optionally tightened by explicit --max-kv-size.
• Report the effective per-slot context through /slots and /health.context_size.
• Update serve --estimate-memory so the preflight uses the same per-slot context and active-sequence count as runtime startup.
• Update CLI help text, docs/environment-variables.md, and CHANGELOG.md.

Note: the issue references stale doc paths (docs/en/..., docs/ko/..., docs/CONTINUOUS_BATCHING.md, docs/man/...) that do not exist in the current tree, so the operator-facing docs were added to the existing environment/flag reference instead.

Validation

• cargo fmt --all -- --check
• cargo clippy -p mlxcel --lib --bin mlxcel --bin mlxcel-server --tests -- -D warnings
• cargo test -p mlxcel --lib context
• cargo test -p mlxcel --lib build_server_config
• cargo test -p mlxcel --bin mlxcel serve_preflight
• git diff --check

Notes

Full cargo test -p mlxcel --lib is still known to fail on the pre-existing NVFP4 sanitize test that was reproduced on origin/main during the preceding issue #52 audit, so this PR uses focused coverage for the changed behavior plus clippy/fmt.