## Motivation

eBPF profiling on a live Hoodi node shows secp256k1 at ~11% of total
CPU, with
53% from P2P transaction validation. `Transaction::sender()` recomputes
ECDSA
public key recovery on every call — when the same transaction has
`sender()`
called multiple times (mempool checks, block building), each triggers a
fresh
recovery.

## Description

Add `sender_cache: OnceCell<Address>` to all transaction types,
following the
same pattern as the existing `inner_hash: OnceCell<H256>` cache for
`Transaction::hash()`. The first `sender()` call computes and caches;
subsequent
calls return the cached value.

Changes:
- Added `sender_cache: OnceCell<Address>` field to all 7 transaction
structs
  (Legacy, EIP2930, EIP1559, EIP4844, EIP7702, PrivilegedL2, FeeToken)
- `sender()` now delegates to `get_or_try_init(compute_sender)` — zero
behavior change
- Field uses `#[rkyv(with=rkyv::with::Skip)]` (same as `inner_hash`)
- All construction sites updated (RLPDecode impls + l1_watcher explicit
construction)
- Sites using `..Default::default()` automatically pick up the new field

## EXPB Benchmark Results

**Fast scenario** (vs main baselines #253/#258):

| Metric | Main | cache-tx-sender | Change |
|--------|------|-----------------|--------|
| latency_avg | 26.23 ms | 25.69 ms | **-2.1%** |
| mgas_avg | 775.78 | 795.35 | **+2.5%** |

**Gigablocks scenario** (vs main #254):

| Metric | Main | cache-tx-sender | Change |
|--------|------|-----------------|--------|
| latency_avg | 754.51 ms | 723.45 ms | **-4.1%** |
| mgas_avg | 1547.14 | 1616.66 | **+4.5%** |

Fast shows ~2% improvement (noise range). Gigablocks shows a meaningful
**4.5%
mgas improvement** — more transactions per block means more cache hits.

## How to Test

```bash
cargo test -p ethrex-common -p ethrex-blockchain -p ethrex-p2p
```