copilot-bluespec: Add more configuration options for floating-point code generation

[RyanGlScott]

Currently, copilot-bluespec has a very specific strategy for generating floating-point operations that is based on Bluespec's FloatingPoint package. This approach has the advantage that it is a 100% native Bluespec implementation, but it also comes with some distinct downsides:

1. The FloatingPoint package lacks implementations of certain floating-point operations (e.g., transcendental functions), which forces us to resort to using C code as shim implementations for these operations. These shims can be simulated, but they cannot be run on actual hardware.
2. The compile times for FloatingPoint operations are quite large due to the complexity of the implementation, combined with how much inlining and pipelining the Bluespec compiler performs. In large examples, using enough floating-point operations can even cause the Bluespec compiler to error out due to the amount of elaboration it has to perform (e.g., see bsc overflows stack space on small Double example B-Lang-org/bsc#1025).
3. Per Missing functions in `FloatingPoint` API? B-Lang-org/bsc#534 (comment), the design of the FloatingPoint package was tailored for writing processors, which may not be as well suited for general numeric computation. For other use cases, it may be desirable to plug in specific floating-point IP libraries written in Verilog.

It is unclear if there is any single approach that will solve all of these problems at once, so we should make copilot-bluespec configurable so as to generate different floating-point-related Bluespec code for different use cases. I can foresee having some combination of the following approaches to pick from:

• The current approach of generating FloatingPoint Bluespec code as-is.
• Generating C shims for all floating-point operations, including basic operations such as addition (+).
• Generate FloatingPoint Bluespec code where possible, but using noinline pragmas (as recommended in bsc overflows stack space on small Double example B-Lang-org/bsc#1025 (comment)) to make the compile times more manageable.
• Delegate some (or all) floating-point operations to Verilog. That is, generate foreign Verilog imports for particular operations on the Bluespec side, and then the user would be responsible for plugging in a floating-point IP package that implements those operations.

I imagine that these options could be folded into BluespecSettings somehow.