Fix Assertion Crash and Optimize Compile Time

[guosran]

Summary

This PR addresses PR #296, and fixes two problems:

1. Assertion crash in cloneWithNewDimensions() when replaying tile/link overrides on smaller CGRA grids
2. Excessive compile time from running the full mapper on every balance probe

Problem 1: Assertion Crash

Root Cause

ResourceAwareTaskOptimizationPass tests task placement on smaller CGRA grids by calling cloneWithNewDimensions(). This method replays all tile_overrides and link_overrides from the original 4×4 grid. However, if an override references coordinates (e.g., (3,3)) that don't exist in the smaller cloned grid, the asserting getTile() call crashes.

Solution

Replaced 4 asserting getTile() calls with safe coord_to_tile_.find() lookups. These changes allow graceful degradation: overrides that don't apply to the new grid's coordinate space are simply skipped.

Problem 2: Excessive Compile Time

Root Cause

The balance phase in ResourceAwareTaskOptimizationPass ran the full mapper on every speculative CGRA count probe.

Solution

Restructured the balance phase to use analytical estimates during iterations.

[tancheng]

This PR is on top of your current register-constraint PR..

[guosran]

This PR is on top of your current register-constraint PR.

Sry my bad, the branch is clean now.

[tancheng]

This PR is on top of your current register-constraint PR.

Sry my bad, the branch is clean now.

looks not yet committed.

[guosran]

This PR is on top of your current register-constraint PR.

Sry my bad, the branch is clean now.

looks not yet committed.

I mistakenly pushed the current pr to the wrong branch yesterday, and I have reverted the changes. This branch is based on ShangkunLi:predicate-handling. Both branches are clean now.

[guosran]

resnet currently takes ~ 3min, other tests finish within 1 sec.