Fix JSONEncoder performance regression in 6.0 toolchain

[kperryua]

Major refactoring of Darwin Foundation's (and hence swift-foundation's) JSONEncoder took place a couple of years ago, with the rewrite being entirely in Swift, no longer relying on NSJSONSerialization. This effort resulted in a significant performance boost on Darwin. However, when swift-corelibs-foundation was re-cored on top of swift-foundation in 6.0, a performance regression was discovered there compared to the 5.10 toolchain.

After examining the code from the 5.10 toolchain's JSONEncoder, I've adapted the swift-foundation implementation to use its general architecture, with some additional important correctness fixes and other adjustments. I also included some other optimizations on top of it. All of this resulted in slightly improved performance in usage within the toolchain (over 5.10), as well via the swift-foundation package and Darwin Foundation.framework over previous.

To summarize the various changes:

• Serialize to a [UInt8] instead of a Data. It's unfortunately noticeably more expensive to serialize to Data right now.
• Avoid unnecessary String allocations when serializing escaped characters in Strings.
• Replace JSONReference (mutable reference / leaf value combination) with JSONEncoderValue (immutable value type) + JSONFuture (container of either a value or one of the reference types) + JSONFuture.RefArray + JSONFuture.RefObject (pure reference types). This approach resulted in less overall retain/release traffic, at the cost of some rare "unwrapping" of JSONEncoderValue values back into the reference types.
• Make each __JSONEncoder responsible for one layer of decoding, instead of reusing the same one multiple times. This enabled some performance optimizations over what the 5.10 SCL-F implementation had, including CodingKey and CodingPath optimizations (each encoder holds a CodingKey, and each nested container has a CodingPathNode linked list descending from the original encoder; the path, when needed (rarely) is constructed by walking up this chain of encoders), and __JSONEncoder instance reuse to avoid allocation churn.
• Reimplemented the _JSONDirectArrayEncodable optimization path to use [UInt8] buffers directly (with substring lengths where needed) instead of holding on to Strings. This will enable us to have the stdlib encode straight to our buffer instead of jumping through an unnecessary transient String allocation if such an API is ever introduced.

[kperryua]

@swift-ci Please test

[kperryua]

Sure!

@jmschonfeld ran our JSON encoding benchmarks for me on Ubuntu, comparing the 5.10, 6.0, and 6.0+this PR versions

Canada-encodeToJSON:

• 5.0: 15 iters/s
• 6.0: 13 iters/s
• This PR: 20 iters/s

Twitter-encodeToJSON:

• 5.0: 549 iters/s
• 6.0: 353 iters/s
• This PR: 658 iters/s