gpl: memoize signal pin counts per cell master during initialization

[oharboe]

Summary

During PlacerBaseCommon::init(), instance pin density calculations previously invoked dbInst::getITerms() and count_signal_pins() for every instance. On large designs (e.g. 2.4 million instances), this caused millions of (N)$ dbNetITermItr::size() linked-list traversals across ODB database structures.

Since an instance's pin count is defined by its LEF cell master (dbMaster), this change memoizes signal pin counts per cell master using a local boost::unordered_flat_map during PlacerBaseCommon::init().

Benefits & Performance

• Reduces getMTerms() evaluations from (\text{instances})$ (e.g., 2.4 million) to (\text{unique masters})$ (e.g., ~200).
• Completely eliminates linked-list getITerms() iteration overhead during GPL initialization without altering any central ODB structures.
• Tested and verified locally with bazelisk test //src/gpl/test:simple01-skip-io-tcl_test //src/gpl/test:gpl_messages_txt_check-py_test.

[gemini-code-assist]

Code Review

This pull request optimizes the average pin density calculation in PlacerBaseCommon::init by memoizing master signal pin counts using a flat map. However, the implementation changes the pin count types from int to double. The review feedback recommends reverting these types back to int to maintain semantic correctness and consistency with the codebase, while ensuring that proper static casts to double are used during division to avoid integer division bugs.

[oharboe]

Applied all Gemini review suggestions in commit b9f6571 (using boost::unordered::unordered_flat_map with discrete int pin counts) and resolved review threads.