Provide a custom tsi_t implementation with loadmem support

[jerryz123]

See ucb-bar/chipyard#683

[colinschmidt]

I'm not sure I understand this implementation. There is upstream support for preloading that was used in the esp-isa-sim version of this. Why can't firesim use that as well?

[colinschmidt]

This is fine.
For reference:
As discussed online it is reinventing the loadmem support from htif_t a bit but using that support would require a rewrite of the firesim_tsi_t implementation that predates the htif_t support. There is no technical reason we can't do this but also a very minimal (if any) technical benefit to doing it. Directly running memcpy might be slightly faster but hard to know since the magic loadmem unit has a large chunk size (@jerryz123 This is true right?) so its only a few function calls per program. It took us a few years to think about changing firesim_fesvr_t so I think this code will live even longer (as it pushed the good enough bar even higher) but 🤷‍♂️

[jerryz123]

The chunk max size appears to be 1024 bytes defined here: https://github.com/riscv/riscv-isa-sim/blob/4baf970f1b151400152c43559e601625dc03bd67/fesvr/tsi.h#L40-L41