fix RUSTC_BACKTRACE always being set
I kept wondering why Miri programs, whenever isolation is disabled, behave as if RUSTC_BACKTRACE was set. Finally I realized it's because some early rustc setup code sets that env var, and that is then propagated to the interpreted program.
So fix that by taking a copy of the environment before any rustc setup, and use that copy as the basis for what is provided to the interpreted program.
implement some missing float functions
With this we support the entire float API surface of the standard library. :)
Also fixes https://github.com/rust-lang/miri/issues/2468 by using host floats to implement FMA.
Some crates are using nightly and failing when mapping these errors,
for example <https://miri.saethlin.dev/?crate=remove_dir_all&version=0.7.0>:
```
error: unsupported operation: io error NotADirectory cannot be translated into a raw os error
--> /root/.rustup/toolchains/miri/lib/rustlib/src/rust/library/std/src/sys/unix/fs.rs:1203:19
```
fix build.rs invoking RUSTC to do check builds
This makes the Miri driver, when invokved via the RUSTC env var from inside a build script, behave almost entirely like rustc. I had to redo how we propagate sysroot information for this (which is actually back to how we used to do sysroot propagation many years ago).
Fixes https://github.com/rust-lang/miri/issues/2431
this requires a change in sysroot handling: miri driver now requires
MIRI_SYSROOT to be set when it is in 'target' mode, rather than relying on
`--sysroot` always being present.
And also expand the cache integrity checks to cover this case, and
generally assert a lot more about the unique_range, then tighten up
sloppy implementation scenarios that this uncovered.
fix some debugging papercuts
- detect when Miri got locally installed and is being run with the wrong toolchain
- when an ICE occurs, print a backtrace of where the interpreter was at the time
