doc: Update the rsz setup repair architecture

[jhkim-pii]

Part 12 of 12 of the PARALLEL_EVAL setup repair stack. It builds on #11804 and also shows the commits of parts 1 to 11; its own change is the last commit.

Summary

• Update src/rsz/doc/architecture.md to the setup repair code of this stack. The table below summarizes the 12 PRs, and the sections after it introduce the new Mt moves, the new options and the PARALLEL_EVAL policy.

Stack

#	PR	Change
1	#11794	LEGACY setup repair path regressions: 16 golden cases on five platforms.
2	#11795	Remove the experimental MT1, MEASURED_VT_SWAP and LEGACY_MT policies; the RSZ_* environment knobs become OptimizationPolicyConfig fields.
3	#11796	Setup repair profiler: -profile_mode 1/2 and -profile_dir write stage times, WNS/TNS samples and events.
4	#11797	Side-path models of the Mt delay estimator: fanin side loads, the opposite transition and the flip-flop checks.
5	#11798	RebufferMt: buffer-tree plans scored by a buffer-tree estimate.
6	#11799	The experimental PARALLEL_EVAL policy with VtSwapMt, SizeUpMt and RebufferMt, and -effort.
7	#11800	SplitLoadMt in the PARALLEL_EVAL default move set.
8	#11801	SwapPinsMt in the PARALLEL_EVAL WNS tail.
9	#11802	UnbufferMt in the PARALLEL_EVAL WNS tail.
10	#11803	SizeDownFanoutMt in the PARALLEL_EVAL WNS tail, and the same steps for every Mt generator.
11	#11804	-profile_mode 3 per-move analysis and report_delay_estimator_accuracy_for_rebuffering.
12	#11805	This PR: the architecture document.

New Mt Moves

• Thread-safe forms of seven serial moves: VtSwapMt, SizeUpMt, RebufferMt, SplitLoadMt, SwapPinsMt, UnbufferMt and SizeDownFanoutMt.
• The main thread reads live STA for each target (prepareInputs()); generate() and estimate() run on worker threads from that data, and each candidate is scored with the Liberty-table DelayEstimator over a window of stages, side paths included.
• The serial moves and the LEGACY policies do not change.

New UI

• repair_timing -phases PARALLEL_EVAL: the experimental policy (warning RSZ-2024).
• repair_timing -effort low|medium|high: how far PARALLEL_EVAL goes after the batch stage stalls; high, the default, runs the WNS tail stage.
• repair_timing -profile_mode 0|1|2|3 and -profile_dir dir: the setup repair profiler, off by default.
• report_delay_estimator_accuracy_for_rebuffering -inst inst [-driver_pin pin] [-delay_levels n]: the buffer-tree estimate of a driver against STA.
• report_delay_estimator_accuracy -estimator takes legacy or stage_window, and the experimental MT1, MEASURED_VT_SWAP and LEGACY_MT phase tokens are removed.

PARALLEL_EVAL Policy

• One RC and STA update per batch instead of per move: worker threads generate and score the candidates of the worst endpoints, and the main thread commits a conflict-free batch in one ECO journal and keeps it only when WNS or TNS improves.
• When the batches stall, the WNS tail stage repairs the worst endpoints with small batches.
• Work is counted in reference units instead of wall time, so a run makes the same decisions on every machine and for any thread count.

[github-actions]

This PR exceeds the recommended size of 1000 lines. Please make sure you are NOT addressing multiple issues with one PR. Note this PR might be rejected due to its size.

[gemini-code-assist]

Code Review

This pull request introduces the experimental PARALLEL_EVAL policy to the Resizer setup repair architecture, enabling parallel evaluation of multiple targets and batch commits to reduce timing updates. It adds several multi-threaded (Mt) move generators and candidates, a setup repair profiler for run-local diagnostics, and a diagnostic command to report rebuffering delay estimator accuracy. Additionally, EstimateParasitics is updated to track observer-estimated nets, allowing timing reads by observers without clearing pending nets. The review feedback highlights several critical safety improvements, primarily recommending defensive null checks in DelayEstimator.cc and Resizer.cc to prevent potential null pointer dereferences, as well as correcting a truncated FNV-1a 64-bit offset basis constant in Rebuffer.cc.

[oharboe]

WNS precent change gives strange numbers because WNS is often small. Perhaps switch to minimum clock period reg2reg change(ignore in/out constraints because they are often very ad-hoc)?

[maliberty]

Here is it called 'keep' but later it is called 'keep on speculation' or just 'Speculate'. One term would be better and speculate is clearer.