Revert "replace unsafe_convert methods with safe cconvert ones"

[Keno]

Reverts #51764. CI is failing on master after that PR.

[vtjnash]

It looks like this is an effect_system / caching issue. Somewhere along the way we compile or run a copy of eltype{::Type{Union{}}) without inferring it. I am trying to find where that happens.

methods(eltype, Tuple{Type{Union{}}})[1].specializations.cache
Core.CodeInstance(MethodInstance for eltype(::Type{Union{}}), 0x0000000000000001, 0xffffffffffffffff, Union{}, #undef, nothing, 0x00000000, 0x00000000, nothing, true, false, 0x00, Ptr{Nothing} @0x00007f74d57fc830, Ptr{Nothing} @0x00007f74d56b1cb0)
Core.CodeInstance(MethodInstance for eltype(::Type{Union{}}), 0x000000000000002d, 0xffffffffffffffff, Union{}, #undef, nothing, 0x000010c0, 0x000010c0, nothing, false, true, 0x01, Ptr{Nothing} @0x0000000000000000, Ptr{Nothing} @0x0000000000000000)

[vtjnash]

Looks like there are two interconnected issues with how effects are hooked up that lead to this fatal result:

1. When typeinf_edge needs effects, it checks the first entry of the cache, even though effects are not guaranteed to be populated (only rettype and world are valid entries in the cache--more on this next)
2. jl_get_method_inferred only ensures that rettype and world are valid to use for IPO, other fields are not permitted to exist or be used by IPO passes, as they are not part of the set of values that get tracked by the compiler when making decisions about how and where to do caching
3. The code returned from typeinf_ext is from the InferenceFrame, not the InferenceResult. When already_inferred is true, then transform_result_for_cache will never be called on this src, resulting in the contents of this object being invalid, as it will never have been populated with the corrected fields BUT the code will have been modified in-place so it should have had the edges and world-bounds fields populated