Fixed incorrect PTX parsing of ret instruction after branch label

[lamarrr]

Description

The PTX parser replaces PTX code with inline PTX code (using inline ASM blocks).
It considers a branch label and the immediate instruction as a single unit to process.
During the ASM->CUDA transform step, it searches for the ret instruction in the string and replaces the whole statement and not the substring that contains the ret; instruction. which means an expression like:

BB0_1:
ret;

is parsed as:

BB0_1: ret;

and then transformed to:

bra RETTGT

instead of:

BB0_1: bra RETTGT

This merge request fixes this bug.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[lamarrr]

thanks for the quick turnaround!

[lamarrr]

/merge