Potential header precedence issue with new cubin/artifact loading approach

[aleozlx]

flashinferMetaInfo.h is now downloaded from artifactory, but if another one is placed inside include/flashinfer/trtllm/batched_gemm/trtllmGen_bmm_export/ (either user has dirty tree or packaging error) it could hide the header downloaded by the new cubin loader approach thereby bypass verification (download is verified but not running the same as the one downloaded) and potentially introduce discrepencies which can be hard to track down (especially when we expand this approach to automate more generated headers in the future).

[aleozlx]

tagging #1328

[aleozlx]

i can't think of a quick fix off the bat, unless we restrict include paths somehow. filing this issue to track first

[yzh119]

Looping in @cyx-6 @joker-eph for their input on this discussion