fix(core): stop injecting color env into test workers - #1893

Motivation

Rstest injected FORCE_COLOR=1 (or NO_COLOR=1 + FORCE_COLOR=0 in agent sessions) into test worker processes so that its own worker-rendered diffs stayed colored. Code under test shares that process.env, so styleText, chalk and picocolors in user code produced ANSI under Rstest but not under plain Node — snapshot tests of CLI output passed locally and failed in CI.

Closes #1774. Related: #1767 / #1775.

Changes

• Rstest never writes FORCE_COLOR / NO_COLOR into any child env. Pool workers and the globalSetup worker are spawned with the host env as-is; config env is applied by setupEnv only.
• The host resolves one boolean for its own output (resolveColorEnabled: NO_COLOR → off, FORCE_COLOR → on unless 0/false, agent session → off, otherwise picocolors detection). Each task gets the same rule applied to its projected env (host + project env), so env: { NO_COLOR: '1' } in config still keeps worker-rendered diff lines plain while host output follows the host env.
• Workers apply the flag to Rstest's own picocolors instance and to the bundled tinyrainbow default instance before renderers load. tinyrainbow is aliased to the copy resolved from @vitest/utils in rslib.config.ts, so the toggle also reaches @vitest/utils diffs and @vitest/expect matcher messages.
• environmentKey keeps the raw FORCE_COLOR / NO_COLOR segment so isolate: false workers are not reused across projects with different color env.
• restart.ts: clearing the screen on restart is gated on TTY only, no longer on color support.

Behavior before / after (no user color env unless stated):

Which switch controls what after this change:

There is no switch that changes Rstest's own colors without also reaching the worker env; a dedicated color option (#1767) is out of scope here.

Downstream repos currently carry workarounds for the injection. Once this lands they can be removed:

Tests: unit table for resolveColorEnabled; e2e for host NO_COLOR / agent / piped output, project NO_COLOR keeping diff lines plain under CI, and isolate: false color affinity. Known untested: the flag turning worker colors on when the worker's own detection says off (needs a pty). No docs change.