[shell-operator] chore: improve json performance

[ldmonster]

Benchmark Results: goccy/go-json vs encoding/json

Marshal (encoding)

Payload	go-json	std json	Speedup	Alloc reduction
Small map (Pod metadata)	1,342 ns	1,708 ns	1.27x	4 vs 13 allocs
Medium map (full Pod)	12,469 ns	16,742 ns	1.34x	25 vs 104 allocs
Binding context (50 objects)	193,023 ns	244,743 ns	1.27x	447 vs 1,611 allocs
Binding context (500 objects)	2,004,233 ns	2,535,465 ns	1.26x	4,452 vs 16,011 allocs
Struct (10 MetricOperations)	6,195 ns	9,233 ns	1.49x	12 vs 52 allocs
AdmissionReview	14,915 ns	19,588 ns	1.31x	29 vs 119 allocs

Unmarshal (decoding)

Payload	go-json	std json	Speedup	Alloc reduction
Small map	1,715 ns	3,118 ns	1.82x	20 vs 25 allocs
Medium map	17,368 ns	23,213 ns	1.34x	187 vs 168 allocs
Binding context (50 objects)	242,985 ns	311,178 ns	1.28x	2,369 vs 2,082 allocs
Binding context (500 objects)	2,512,677 ns	3,141,470 ns	1.25x	23,522 vs 20,535 allocs
Struct (10 MetricOperations)	10,211 ns	31,216 ns	3.06x	83 vs 132 allocs
AdmissionReview	20,201 ns	32,358 ns	1.60x	203 vs 199 allocs

Encoder/Decoder (streaming)

Payload	go-json	std json	Speedup
Encoder (500 obj binding ctx)	1,802,357 ns	2,402,493 ns	1.33x
Decoder (20 MetricOperations)	25,570 ns	66,858 ns	2.61x

Key takeaways

• Struct unmarshal (used for MetricOperations, AdmissionReview parsing) sees the largest gain: 2.6-3.1x faster
• Map marshal (used for binding contexts, checksum computation) is consistently 1.25-1.34x faster with 3.6x fewer allocations
• Streaming decode (used for MetricOperationsFromReader, webhook request parsing) is 2.6x faster
• go-json trades slightly more bytes-per-alloc for dramatically fewer allocation count on the marshal side, reducing GC pressure