fix(dual-franka): align deployment, interaction and task prompts

[NewYork424]

This PR consolidates fixes developed during dual-Franka physical demo reproduction and targets main. It builds on the native tool registration from #198, which is included in main.

• Use official RLinf environment, controller, gripper and camera APIs through an RPent-owned adapter, with no RLinf source patches. Preserve policy observations while retaining uncropped RGB-D and camera metadata from the same acquisition.
• Refresh observations after operator interaction, reject stale localization inputs, retain gripper state during analytic motion and recovery, and hold TCP position while rotating. Keep deployment-specific controller tuning in configuration.
• Support operator-confirmed reset and verdicts through the terminal or the existing Dashboard composer. Preserve feedback across attempts and retire cancelled, disconnected or timed-out work before accepting new actions. Recognize reset aborts with notes and return validation errors for malformed Dashboard operator context. The reviewed terminal memory opt-out issue remains unchanged: direct /success can still publish memory despite --no-auto-merge-memory.
• Keep manual skill schemas derived from native declarations, expose VLA deployment metadata, and support calibration sample resume/exclusion.
• Keep scene-specific camera and manipulation strategies in task definitions rather than the shared robot prompt. Update Task 5 so the spoon goes into the upright cup, register Task 6 for two-cup dirty/clean sorting without utensils, and preserve each task's VLA boundaries and operator-verdict requirements.

[NewYork424]

Superseded by the following focused pull requests, in review order:

• fix(dual-franka): align RLinf v0.4 deployment and control diagnostics #252 — fix(dual-franka): adapt deployment to official RLinf v0.4
• fix(dual-franka): hold TCP position during in-place rotation #253 — fix(dual-franka): hold TCP position during in-place rotation
• feat(dual-franka): add manual continuous VLA diagnostics #254 — feat(dual-franka): add manual continuous VLA diagnostics
• fix(interaction): preserve attended runs and current observations #255 — fix(dual-franka): refresh current observations before localization
• fix(interaction): preserve attended attempts and feedback across replies #256 — fix(interaction): preserve attended attempts and feedback across replies
• feat(dashboard): support attended physical task confirmations #257 — feat(dashboard): support attended physical task confirmations
• feat(calibration): resume collection and exclude invalid samples #258 — feat(calibration): resume collection and exclude invalid samples
• refactor(dual-franka): separate task policies from robot prompts #259 — refactor(dual-franka): separate task policies from robot prompts

Dependencies: #253 and #254 depend on #252; #256 depends on #255; #257 depends on #256. Dependent PRs target their prerequisite branches to keep diffs focused and should be retargeted to main after those prerequisites merge. The default Task6 in #254 is supplied by #259; until then, select an existing task or an explicit instruction.

Closing this combined PR in favor of the split series.