fix(cli): reject unparseable numeric flags in benchmark and chat

[addyCooks]

Description

nanotune train validates its numeric flags before it uses them, but benchmark and chat parsed the identical class of flag with a bare Number.parseInt/parseFloat and no check at all.

parseInt stops at the first character it cannot parse, so --ctx-size 4096x silently became 4096 and the run proceeded under a value the user never typed. Anything that failed outright became NaN and reached llama-server unchecked: String(NaN) is "NaN", so --gpu-layers abc spawned the real argument -ngl NaN, while JSON.stringify(NaN) is null, so a bad --max-tokens/--temperature/--top-p left the field out of the completion body and llama-server quietly used its own default. --timeout was worse still setTimeout(fn, NaN) fires immediately, so a typo aborted every test the instant it started.

Every numeric flag on both commands now goes through one shared parseNumericFlag() in chat-helpers.ts, which uses Number() rather than parseInt/parseFloat so trailing garbage fails instead of truncating, and rejects a fractional value where llama-server wants an integer. A rejected flag is reported as Invalid value for --flag, matching train's existing message, and the command stops before any subprocess or network work: for chat ahead of the project check, for benchmark alongside the sampling flags, ahead of the base-model download and the server spawn.

benchmark's flag wiring moves into resolveBenchmarkFlags() in benchmark-utils.ts, next to the sampling flags it already resolved there, which also drops the duplicate copy of chat's server-option builder and lets the whole path be unit-tested. Flags a --preset overrides are still parsed, so a typo alongside a preset is reported rather than silently discarded.

Closes #134.

Behavior changes

• --ctx-size 4096.5 is now rejected rather than truncated to 4096.
• A typo'd numeric flag passed alongside a valid --preset now errors instead of being ignored (the preset still wins on values).

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update

Testing

Automated Tests

• All existing tests pass (pnpm test:all completes successfully)
• New tests added for new functionality (if applicable)

Manual Testing

• Tested nanotune init
• Tested nanotune data commands (add/import/list/validate)
• Tested nanotune train
• Tested nanotune export
• Tested nanotune benchmark

Checklist

• Code follows project style guidelines (pnpm format)
• Self-review completed
• Documentation updated (if needed)
• No breaking changes (or clearly documented)