Optimize compile time of Deserializer trait methods

[dtolnay]

We can instantiate the control flow in these methods once per Read type rather than once per Visitor type. This reduces compile time of https://github.com/serde-rs/json-benchmark by 44% as measured by cargo clean && RUSTFLAGS='-C codegen-units=1' cargo build --release --no-default-features --features parse-struct,lib-serde,all-files -Ztimings, with 0-13% impact on performance.

Before

   Completed serde_json v1.0.55 in 3.8s
   Completed json-benchmark v0.0.1 bin "json-benchmark" in 18.1s

                                DOM                  STRUCT
======= serde_json ======= parse|stringify ===== parse|stringify ====
data/canada.json         270 MB/s              620 MB/s
data/citm_catalog.json   390 MB/s              950 MB/s
data/twitter.json        270 MB/s              560 MB/s

After

   Completed serde_json v1.0.55 in 3.7s
   Completed json-benchmark v0.0.1 bin "json-benchmark" in 10.1s

                                DOM                  STRUCT
======= serde_json ======= parse|stringify ===== parse|stringify ====
data/canada.json         250 MB/s              540 MB/s
data/citm_catalog.json   380 MB/s              850 MB/s
data/twitter.json        260 MB/s              560 MB/s

[Marwes]

This could be done in a safe manner though I guess there would be some additional branches unless the unreachable!() branches are made unreachable_unchecked. https://github.com/Marwes/json/tree/safe_once_dispatch

It is only a POC but I'd expect the vtable from the trait to end up the same as this unsafe impl.

[Marwes]

cc #313