rcx: fix bench_wires segfault on duplicate pattern names

[gadfort]

Reopens #7897 — the merged fix 788fca19f1 is narrower than the bug.

Symptom

bench_wires -all on a technology with thick top metal and a coarse database
unit emits a duplicate-net warning and then dies:

[INFO RCX-0055] Finished 72 bench measurements for pattern MET_OVER_MET
[WARNING RCX-0409] Cannot create net R6_M3oM0_W2200W2200_S00000S01800_3, duplicate net
Signal 11 received

Symbolized:

odb::dbNet::get1stBTerm()
rcx::extMeasure::createNetSingleWire()
rcx::extRCModel::writeBenchWires_DB_res()
rcx::extRCModel::benchDB_WS()
rcx::extRCModel::linesOverBench()

Root cause

The pattern name is doubly quantized before it reaches the net name:

• setTargetParams floors the coupling spacing to a 10nm grid
  (n2 = (n2 / 10) * 10)
• mkNet_prefix then rescales by (1000 * s_nm) / dbunit with integer
  division

So two entries of the spacing table built in benchDB_WS collide whenever they
land in the same bucket, even though they are distinct as doubles.

Concretely, on a layer with WIDTH 2.200 / SPACING 1.800 / PITCH 4.000 and
DATABASE MICRONS 2000, the table holds both SPACING and PITCH - WIDTH.
Those are equal in exact arithmetic, differ by exactly one ULP as doubles, and
both round to 1800 in the name.

788fca19f1 guarded this with std::fabs(spacing - s1) > DBL_EPSILON. That is
insufficient in two ways:

1. It guards only that one hardcoded pair. The s1 and s2 entries added by
   the later loop iterations are never compared against anything.
2. DBL_EPSILON is an absolute tolerance, meaningful only for values near
   1.0. One ULP of 3.6 is 4.44e-16, which is > DBL_EPSILON, so the guard
   does not fire and both values enter the table.

More fundamentally, because the name is quantized to 10nm, two spacings that
are genuinely several nm apart can also collide. No floating point tolerance
can catch that.

Fix

• Dedupe on the generated name. benchDB_WS now skips any
  (width, spacing, spacing2) combination whose pattern name was already
  emitted, keyed on _wireDirName itself. That is the value that actually has
  to be unique, so it covers both collision mechanisms and any future one. The
  set is local to one benchDB_WS call, which is correct — the name embeds
  the pattern kind and the metal/under-metal pair, so those cannot alias
  across calls.
• Make a null net a hard error. All three extMeasure::create*NetSingleWire*
  wrappers dereferenced the null that dbCreateNetUtil::createNetSingleWire
  returns. They now report RCX-529 / RCX-530 / RCX-531. Reaching this is
  a bug, so it should fail loudly rather than crash or silently drop wires.
• Give the solver flow a logger. The four extRCModel::lines* pattern
  generators built their extMeasure with a null logger, and one of them
  reaches createDiagNetSingleWire. Passing the model logger keeps the new
  error reportable instead of turning into another null dereference.

The DBL_EPSILON guard is left in place. It is now redundant, but removing it
would add entries to the spacing table and change pattern output for every
technology where it currently fires — out of scope for a crash fix.

Testing

New regression src/rcx/test/dup_pattern_names, registered in both CMake and
Bazel. It ships a minimal hand-written routing-only tech LEF (no vendor file)
carrying the geometry that triggers the collision, and asserts the colliding
pattern is generated exactly once with all of its wires. Before this change the
test reproduces the reported crash with the identical net name.

bazel test //src/rcx/test/... — 18/18 pass. generate_pattern is unchanged,
confirming the dedupe does not over-fire on technologies that were already
fine.

Type of Change

• Bug fix

Impact

No segfault on bench creation

Verification

• I have verified that the local build succeeds (./etc/Build.sh).
• I have run the relevant tests and they pass.
• My code follows the repository's formatting guidelines.
• I have included tests to prevent regressions.
• I have signed my commits (DCO).

Related Issues

#7897

[gemini-code-assist]

Code Review

This pull request addresses a crash caused by duplicate pattern names during bench wire generation by tracking generated pattern names in a set and skipping duplicates. It also updates extMeasure to use the proper logger and adds error logging when net creation fails. The review feedback highlights critical safety issues where the code still dereferences potentially null net pointers immediately after logging the error, which would lead to crashes. Additionally, the feedback recommends adding null checks for _wireDirName and logger_ to prevent undefined behavior and crashes in legacy execution paths.