[llvm] [MC][AsmParser]: Cache files loaded by .incbin across repeated directives (PR #194254)

[via llvm-commits]

SergeantCooper wrote:

@aengelke I haven't tested this with wine but have tested with a minimal producer which I believe replicates the erroneous  behavior related to wine's source code
The bibisected [commit](https://gitlab.winehq.org/wine/wine/-/commit/0b3f90ab1485d5bd32bd72d41c7fd8213b3b95b9)  has this change
```
for (i = 0, res = spec->resources; i < spec->nb_resources; i++, res++)
{
    output( "\n\t.align %d\n", get_alignment(4) );
    output( ".L__wine_spec_res_%d:\n", i );
    output( "\t.incbin \"%s\",%d,%d\n", res->input_name, res->input_offset, res->data_size );
}
```

I used the following assembly file for testing
```
.data
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
.incbin "blob.bin", 0, 16
... till 1000 reps
```
The blob size was 1M and generated an output file with both system and local build llvm-mc for which the memory details are mentioned in the description.

https://github.com/llvm/llvm-project/pull/194254