[DjDeveloperr]

Three follow-ups to #505 and #508, from maralcbr's review of the omacom mirror (omacom/linux-aurora#244). The same commits are on that PR.

1. take a G17 render-only kill's cookie before its admission hold. If the cookie counter was exhausted, the render-only kill returned EOVERFLOW with its admission hold still installed. The cookie now comes first, as in a failed context's kill.
2. keep a G17 kept object's list reservation with the object. When reserve_kept failed, keep was still set, so the drop decremented and could consume another kept object's reservation. reserve_kept now reports whether it reserved; an object kept without a reservation is leaked at drop, as documented.
3. count kept G17 render graphs and fix two stale comments.
   • The capacity view reports firmware_linked_render_graphs kept= leaked= beside firmware_linked_objects.
   • Render observation needs only the fragment notification (drm/asahi: halve G17 firmware notifications per render (first-kick word, retire on fragment) #504).
   • A kill hold ends at the consumed scheduler release (drm/asahi: kill faulted G17 contexts before freeing their render work #505).

Testing

Neo (A18 Pro), Mesa agx3 rb6, module built from this branch (build 2c8d9952):

• regression gate: PERFORMANCE-REGRESSION-PASS;
• targeted CTS (5368 cases): 4687 pass, 681 not supported, 0 fail;
• no asahi kernel errors. The capacity view shows the new line, with 0 kept and 0 leaked, and nothing pending.

The omacom mirror passes omacom KUnit with these commits.