perf: optimize identity function composition

[He-Pin]

Motivation

bench/resources/cpp_suite/bench.07.jsonnet builds a lazy array of composed functions:

local f2(f) = function(x) f(f(x));
local id(x) = x;
local slowId = std.makeArray(20, function(i) if i == 0 then id else f2(slowId[i - 1]));
slowId[15](42)

On upstream master, Scala Native overflows the stack for this case even with --max-stack 100000. On JVM, the same identity-equivalent chain still materializes a deep nested function/lazy evaluation path.

Modification

• Add a unary identity fast path in Val.Func.apply1.
• Detect the exact ordinary, non-tailstrict function(x) f(f(x)) shape and return a lightweight identity-composition wrapper.
• Keep the identity-chain probe iterative, so it is friendly to Scala Native stacks.
• Use an allocation-free byte state machine for the base-identity probe: BaseIdentityProbeInProgress marks active probes, while BaseIdentityKnownIdentity / BaseIdentityKnownNonIdentity cache the final predicate result. Recursive composition cycles are treated as non-identity for this optimization and fall back to normal application/max-stack handling.
• Preserve laziness and tailstrict semantics: constructing f2(error "...") stays lazy, calling it still forces the original error, and f2(error "...") tailstrict still forces eagerly.

Result

Compared revisions:

• upstream/master: 8b67cb1
• This PR: 515691b3
• jrsonnet: 5b43fa8

JMH, bench/resources/cpp_suite/bench.07.jsonnet:

./mill -i bench.runRegressions bench/resources/cpp_suite/bench.07.jsonnet

Target	Result
upstream/master	3.013 ms/op
this PR	0.037 ms/op
delta	81x faster

Scala Native CLI, same input:

hyperfine --shell=none --warmup 10 --runs 100 \
  --command-name 'sjsonnet native #826' \
  './out/sjsonnet/native/3.3.7/nativeLink.dest/out --max-stack 100000 bench/resources/cpp_suite/bench.07.jsonnet' \
  --command-name 'jrsonnet 5b43fa8' \
  '/path/to/jrsonnet -s 100000 bench/resources/cpp_suite/bench.07.jsonnet'

Target	Result
upstream/master native	StackOverflowError
this PR native	4.2 +/- 0.9 ms
jrsonnet 5b43fa8	13.1 +/- 1.3 ms
this PR vs jrsonnet	3.09x faster

The native CLI case is tiny and includes process startup, so the stronger signal is the JMH delta plus the master Native failure/pass change.

Verification

Local verification after the semantic state-name update:

git diff --check
./mill -i 'sjsonnet.jvm[3.3.7].test.testOnly' sjsonnet.EvaluatorTests
./mill -i 'sjsonnet.native[3.3.7].test.testOnly' sjsonnet.EvaluatorTests
./mill -i bench.runRegressions bench/resources/cpp_suite/bench.07.jsonnet
./mill -i 'sjsonnet.native[3.3.7].nativeLink'
./mill -i __.checkFormat

All passed.

Boundary Checks

• The optimization only matches the exact ordinary apply shape function(x) f(f(x)).
• tailstrict calls are not folded into the lazy identity-composition path.
• Call-time forcing still preserves the original lazy error: local g = f2(error "..."); g(1).
• Non-function bases still report the normal type error: f2(1)(1).
• Non-identity composed functions still fall back to normal evaluation; covered by f2(function(x) x + 1)(1) == 3 and a multi-layer non-identity chain.
• Recursive identity-composition cycles are covered by self-recursive and object-mutual-recursive tests; they now hit the normal max-stack path instead of hanging in the optimization probe.
• TailCall-related implementation changes are intentionally not included in this PR.

[He-Pin]

Safety update pushed in 53556cfd:

• Added an IdentityChecking state so identity-composition probing detects recursive composition cycles and falls back to normal application instead of spinning in the optimization path.
• Added self-recursive and object-mutual-recursive regression tests.

Validation:

• ./mill -i sjsonnet.jvm[3.3.7].test.testOnly sjsonnet.EvaluatorTests
• ./mill -i sjsonnet.native[3.3.7].test.testOnly sjsonnet.EvaluatorTests
• ./mill -i bench.runRegressions bench/resources/cpp_suite/bench.07.jsonnet -> 0.039 ms/op
• ./mill -i __.checkFormat