feature: Opportunistic drain balancing

[seanjnkns]

Introduces opportunistic balancing of pgs during Drain

From testing, and also having 124 remapped pgs due to osd.23 reweight values temporarily adjusted, we have the following pg -> osd spread:

Projected PG load spread (drain off OSD 0 → OSDs 23–46)

Comparison

	Reservation-only (prior behavior)	PG count + reservations + tie-break (current)
Final min / max	158 / 193	185 / 188
Final spread	35	3

Per-OSD projected Final (new algorithm)

OSD	Baseline	+Upmap	Final
23	158	27	185
24	164	23	187
25	171	14	185
26	173	13	186
27	174	13	187
28	173	13	186
29	173	13	186
30	172	13	185
31	172	15	187
32	175	12	187
33	173	13	186
34	173	13	186
35	173	13	186
36	174	12	186
37	174	12	186
38	173	13	186
39	173	13	186
40	173	14	187
41	173	13	186
42	173	13	186
43	173	13	186
44	173	13	186
45	173	15	188
46	174	13	187

Tie-breaking can swap a single PG between targets with identical rank (e.g. OSDs 32 vs 35); spread stays in the same band.

---

Prior behavior — per-OSD Final (reservation-only, from old algorithm)

OSD	Baseline	+Upmap	Final
23	158	0	158
24	164	13	177
25	171	17	188
26	173	18	191
27	174	13	187
28	173	17	190
29	173	18	191
30	172	13	185
31	172	17	189
32	175	18	193
33	173	13	186
34	173	17	190
35	173	18	191
36	174	13	187
37	174	17	191
38	173	18	191
39	173	13	186
40	173	17	190
41	173	18	191
42	173	13	186
43	173	17	190
44	173	18	191
45	173	13	186
46	174	17	191

---

[seanjnkns]

This algorithm change will also benefit undo-umap, achieving a more balanced distribution.

[Matt1360]

To be clear, this does change the priority of the drain command from concurrency focus to balance focus, which is a tradeoff that I don't think matters for our use case today since we can control that better if we need to. This effectively removes the need for data to move twice, albeit slower, which is likely a win to most.