Fix uppercase cl100k contraction matching

[MikeeI]

Summary

• correct the generated cl100k_base contraction set from the widened [R-Tr-tſ] range to [STstſ]
• add a regression case requiring uppercase "'RE" to encode as token 95253

Why

The precompiled regexp2cg engine currently treats R/r as part of the case-insensitive s|t alternative. Because that branch is atomic and precedes the re alternative, "'RE" is split into "'R" and "E", producing two tokens instead of the OpenAI-compatible single token. The source regex and vocabulary are already correct; the widened generated ASCII bitmap is the divergence.

Both the pinned regexp2cg v0.7.2 and current v0.9.0 regenerate the widened range, so this is a localized correction to the committed generated engine rather than a dependency bump.

Verification

• go test -count=1 ./...
• complete differential check of 10,000 ordinary-text cases against Python tiktoken 0.13.0 across r50k_base, p50k_base, cl100k_base, and o200k_base

Closes #24

[MikeeI]

I opened an upstream regexp2cg report with the minimal interpreter-vs-generated-engine reproduction and bitmap analysis: dlclark/regexp2cg#5

[bluescreen10]

Thanks for this report and more importantly for opening up an upstream report. Let me see how fast this gets attended upstream. if not I can merge the fix.