feat(cli): --safe-eval first-run sandbox profile (#60)

[quangdang46]

What

Adds the safe-evaluation profile requested in issue #60. A single flag (or env var) layers a conservative sandbox on top of any other flags so users can evaluate jcode before pointing it at their main machine, primary credentials, or sensitive repositories.

Usage

jcode --safe-eval run "say hello"
JCODE_SAFE_EVAL=1 jcode run "..."

What it does

Translated at startup into env overrides (each one already honored by existing code paths in master, so this PR is a thin coordination layer + banner + docs, not a parallel reimplementation):

Env var	Effect	Honored by
JCODE_HOME=~/.jcode-safe-eval/ (only if not already set)	Isolated config / sessions / memory / auth dir	storage::jcode_dir
JCODE_OFFLINE=1	Disables update check, telemetry, model-list refresh	Issue #24 / PR #197
JCODE_NO_TELEMETRY=1	Belt-and-suspenders telemetry off	telemetry::is_enabled
JCODE_AMBIENT_DISABLED=1	Ambient runner does not start	ambient runner gate
JCODE_NO_SELFDEV=1	Self-dev auto-detection suppressed	selfdev autodetect

A short banner is printed at startup so users can confirm the profile took effect (suppressed by --quiet):

Safe-eval profile: isolated JCODE_HOME, telemetry off, offline, ambient/selfdev gated.
  JCODE_HOME = /home/<user>/.jcode-safe-eval

Changes

• src/cli/args.rs: --safe-eval global flag.
• src/cli/startup.rs: parse_and_prepare_args translates the flag (and pre-existing JCODE_SAFE_EVAL=1 env) into the 5 coordinated env vars.
• docs/SAFE_EVALUATION.md (new): walks through what's disabled, what isn't, the recommended first-run workflow, cleanup (rm -rf ~/.jcode-safe-eval), and how to compose with other flags.

Tests

cargo check -p jcode --lib clean. The behavior is purely env-var translation; the downstream effects each have their own existing tests (see update::tests::check_for_update_blocking_returns_none_in_offline_mode etc.).

Closes #60