Zero copy deserialization to reduce GC pressure

[eclairss17]

During block verification DeserializeProof, creates a copy of each stem which creates high volume GC pressure during high throughput scale. We introduce a shared view of reference buffer that stores the deserialized proof in its wire format, along with a counter to track number of ProofViews reading from it without having to copy each stem.

Benchmark Results (Apple M1 Pro, Go 1.23, go test -benchmem -benchtime=5s)

BenchmarkCopying_10K              566,760 B/op   10,001 allocs/op   146 µs/op
BenchmarkZeroCopy_10K             242,004 B/op   10,003 allocs/op   242 µs/op
BenchmarkZeroCopy_AccessOnly           42 B/op        1 alloc/op     3.4 µs/op
BenchmarkZeroCopy_SingleProofView       0 B/op        0 allocs/op   14.2 ns/op

GC Memory Overhead Results

Heap allocation reduction per node:

Copying:    566 KB × 7,200 blocks/day  = ~3.97 GB allocated and freed per day
                                        = ~1.45 TB per year per node
Zero-copy:  242 KB × 7,200 blocks/day  = ~1.70 GB allocated and freed per day
                                        = ~0.62 TB per year per node
Savings:    ~2.27 GB/day per node      = ~829 GB/year per node

Aggregate across the network (6,000 nodes):

Per day:    2.27 GB × 6,000 nodes  = ~13.6 TB of heap churn eliminated per day
Per year:   829 GB × 6,000 nodes   = ~4.85 PB of heap churn eliminated per year