Fix comparison of RawVals of differing types and objectness

[brson]

Edit: I have posted a prerequisite to this patch in #773

What

Make comparison of RawVals agree with comparison of ScVals in the case where one value is an object and one value is a smallval, and they are not the same ScValType.

Also fix an unsoundness in RawVal::get_tag due to incorrect assumptions in an unsafe cast.

Why

Per #743, comparison operations for RawVals and ScVals must agree.

The final fallthrough case of Host::obj_cmp does a comparison based on the Tag discriminant, and the Tag discriminants do not have the same ordering as the ScVal discriminants. This patch adds a method Tag::get_scval_type, and changes obj_cmp to use that instead.

This adds a thorough test case which does an exhaustive conversion between u8s and tags, and an exhaustive match on Tags. The exhaustive match triggered a segfault due to an unsound cast from u8 to Tag in RawVal::get_tag. That function assumed that Tag's discriminants were contiguous between the "marker" discriminants ObjectCodeLowerBound and ObjectCodeUpperBound, which was not true. get_tag looks like

    #[inline(always)]
    pub const fn get_tag(self) -> Tag {
        let tag = self.get_tag_u8();
        const A: u8 = Tag::SmallCodeUpperBound as u8;
        const B: u8 = Tag::ObjectCodeLowerBound as u8;
        const C: u8 = Tag::ObjectCodeUpperBound as u8;
        if !((tag < A) || (B < tag && tag < C)) {
            return Tag::Bad;
        }
        // Transmuting an integer to an enum is UB if outside the defined enum
        // value set, so we need to test above to be safe. Note that it's ok for
        // `get_tag` here to be a _little_ slow since it's not called in a lot
        // of small/tight paths, only when doing a switch-based comparison. Most
        // small paths call `has_tag` which tests a _known_ enum case against
        // the tag byte, and therefore doesn't need the range check.
        unsafe { ::core::mem::transmute(tag) }
    }

In safe code it may not be possible to trigger the unsoundness - it may require unsafely constructing RawVal. It is not clear to me that contracts can find this code path, but it's probably best to fix it.

I changed the discriminants of Tag to make RawVal::get_tag always correct, and added docs, but the code is brittle to changes. I considered how to make it less error prone and add tests, but could not come up with a better test case than the one that is already in this patch.

This adds a dev-dependency on itertools for the cartesian_product method.

Known limitations

na

[brson]

Actually the test here is not exhaustive in the conversion of u8s to Tags. There could be a better test of RawVal::get_tag, but I'll wait for some review before thinking about it more.

[brson]

I can reproduce the CI failures. Seems like I've broken something.

[brson]

The CI failure was because changing Tag discriminants is ABI-breaking and requires rebuilding the soroban-test-wasms. I tried, but the wasm-workspace does not build presently. Instead I wrote a testcase and simple, possibly temporary, fix and posted a separate pr #773

[brson]

Rebased. Still needs #767 to land first.

[graydon]

I think this needs to be panic-free

[brson]

I removed the panics from Tag::get_scval_type. I also realized that I had other comparison tests located in another module, so I moved the tests from this PR there.

[brson]

Also did a rebase to pick up the commits from #773

[graydon]

LGTM, thanks!