fix(codegen): bound instruction modifier searches to the current token

[paulzhng]

What this fixes

Feature detection repeatedly searched the rest of the LLVM module for escaped whitespace. Large modules therefore spent quadratic time scanning ordinary instruction tokens.

before: each token searches the remaining module
 after: each token searches only up to its first ordinary delimiter

The earliest delimiter is unchanged, including escaped and Unicode whitespace. Architecture and PTX requirements keep the same meaning.

Verification

• All 1,261 codegen tests passed on the current-main rebase, including terminal LLVM/PTX checks.
• An independent harness compared the exact old/new functions on 6,084 delimiter and near-match cases with identical results.
• On a synthetic 32,768-token module, the old scan took 15.0 seconds and the new scan 1.81 milliseconds. This is a local scanner measurement, not an end-to-end build-speed claim.
• Strict Clippy, formatting and warning-free docs passed. All seven hosted workflows passed at 4953e1e1.

No issue is closed.

[nihalpasham]

I rebased this and checked the token-boundary behavior independently. All 6,084 delimiter/near-match cases agreed with the old implementation, and the large-module scan now scales linearly in the measured cases.

All 1,261 codegen tests, strict Clippy and docs pass. All seven hosted workflows passed at 4953e1e1.

No issue is closed.