shared: switch to structured error handling

[BurntSushi]

This completes the work started in #453 by switching everything in
jiff::shared to structured error handling. This module is somewhat
special in that it is "shared" with jiff-static.

As a result, jiff::shared had its own special bare bones error
type that didn't build a dynamic chain internally (which is what
jiff::Error does, similar to anyhow::Error). It has now been
replaced with structured errors entirely and no dynamic chains. I did
this because I didn't feel like rebuilding the chaining functionality
(or moving it into shared), and because it didn't seem that onerous
of a task.

But doing it this way did make it clear to me that the chaining in
jiff::Error is letting us avoid a lot of tedium. Without the dynamic
chaining, you wind up needing to define many more error types in order
to preserve decent contextualized error messages. It's quite annoying,
but doable at a small scale.

This does shrink LLVM lines in my Biff benchmark even more. Before this
PR, we were at 756,177 lines. After this PR, we are at 753,960. Nice!

In terms of compile times, it seems there isn't much of a change. Before
this PR:

Benchmark 1: cargo b
  Time (mean ± σ):      4.542 s ±  0.017 s    [User: 15.310 s, System: 2.128 s]
  Range (min … max):    4.518 s …  4.570 s    10 runs

Benchmark 2: cargo b -r
  Time (mean ± σ):      7.708 s ±  0.051 s    [User: 65.504 s, System: 2.580 s]
  Range (min … max):    7.633 s …  7.765 s    10 runs

And after:

Benchmark 1: cargo b
  Time (mean ± σ):      4.613 s ±  0.014 s    [User: 15.509 s, System: 2.116 s]
  Range (min … max):    4.595 s …  4.636 s    10 runs

Benchmark 2: cargo b -r
  Time (mean ± σ):      7.767 s ±  0.066 s    [User: 65.627 s, System: 2.588 s]
  Range (min … max):    7.689 s …  7.862 s    10 runs

Which is somewhat disappointing.