Fix HiCache transfer timing and enable_metrics threading

[ishandhanani]

Follow-up to sgl-project#26974 — re-tested the updated branch on 1x L40S (Qwen3-0.6B, HiCache, write_through, page_size 64, --enable-metrics --log-level debug). All previously-flagged issues are fixed: the decode-side ack consumers work, write-back paths record, the redundant counters are gone, and counters are exact (bytes/token = 114688 = 2 x 28 layers x 8 KV heads x 128 dim x 2B for Qwen3-0.6B). Two remaining issues found, fixed here.

1. Transfer durations were always host wall-clock, never device time

Every event pair fed to elapsed_time() is created without enable_timing=True, so on CUDA elapsed_time() raises

ValueError: Both events must be created with argument 'enable_timing=True'.

and _transfer_elapsed_us silently falls into its except fallback on every transfer. The fallback measures enqueue -> ack-poll latency, not DMA time: an idle onboard of 161 MB logged xfer_us=5751205 (5.75 s, 0.03 GB/s) for a ~7 ms transfer, because the ack sat in ack_load_queue until the next request triggered loading_check. The "prefer device timing" path was dead code.

Fix: create the pairs actually used for timing with enable_timing=True (write start/finish in both controllers, LayerLoadingEvent.start_event + last layer event, which doubles as finish_event). The host fallback remains for backends without elapsed_time().

2. enable_metrics only reached HiRadixCache's controller

decode_kvcache_offload_manager.py and all five HybridCacheController construction sites in hybrid_pool_assembler.py left enable_metrics defaulted to False, so the disagg-decode / hybrid / unified stacks emitted DEBUG logs but never Prometheus, even with --enable-metrics. Threaded enable_metrics + extra_metric_labels through.

Also reverts the unrelated docs_new/index.mdx blog-card rotation (same change already removed from sibling PR sgl-project#26976).

Verification (same box/config)

Check	Before fix	After fix
Idle onboard duration (161 MB load-back, ack polled seconds later)	xfer_us=5751205 (5.75 s, 0.03 GB/s)	xfer_us=6781 (6.8 ms, 25.98 GB/s)
Offload per-op duration	6.5-7 ms (wall-clock incl. poll latency)	4.8 ms consistent, ~25.9 GB/s
Cumulative bandwidth under aiperf load (c=16, 256 reqs)	offload 1.97 GB/s	offload 9.81 GB/s, onboard 24.92 GB/s
DEBUG cumulative vs Prometheus _count/_sum	equal	equal (542 ev / 2832 blk / 20786970624 B / 2118621 us, both channels)
Unified stack Prometheus (SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 + HiCache)	0 series (collector never created)	full series, exact counts (offload 160 blocks, onboard 24 blocks for a 1536-token load-back)
Disagg ack mock tests (test_specv2_kvcache_offloading.py)	13 passed	13 passed
Unified radix cache load/load-back unit tests	58 passed	58 passed
Gating without --enable-metrics (INFO level, load-back traffic)	0 DEBUG lines, 0 series	0 DEBUG lines, 0 series
Block/byte counters	exact	exact (unchanged)

hi_mamba / SWA hybrid stacks need hybrid models this box cannot run; the threading change there is identical to the unified path verified above.