fix(gas-manager): keep legacy gas price history per chain

[fichiokaku]

Problem

GasManagerService.getGasPrice maintained a single rolling history shared by every chain the node serves:

private legacyGasPriceHistory: bigint[] = [];   // one array for all chains

Gas prices across our chains differ by orders of magnitude, so the median of that mixture is meaningless. The outlier checks then throw away a chain's real price in favour of the mixture's median:

} else if (currentToMedianRatio > 500n) {
  estimatedPrice = medianGasPrice;   // "outlier"
} else if (currentToMedianRatio < 50n) {
  estimatedPrice = medianGasPrice;   // "stale/broken"
}

The result is that a chain gets priced using an unrelated chain's gas price.

Impact observed on live nodes

Staging — Chiliz Spicy dominates traffic (2,925 log lines vs 18 for Fuji in one window), so it sets the median:

Chain	Real	Node used
Avalanche Fuji	160 wei	2,751,100,000,000 wei

That value is exactly the Chiliz median plus the 10% standard buffer (2501 gwei x 1.10). A client hit this as repeated execution failures on Fuji.

Production — the cheap chains dominate instead, so expensive chains are badly underpriced:

Chain	Real	Node used	Off by
Chiliz	2,501 gwei	0.0387 gwei	64,625x
Polygon	281 gwei	0.0698 gwei	4,028x
Monad	102 gwei	0.0387 gwei	2,636x
Apechain	102 gwei	0.0550 gwei	1,849x
Sonic / Sei	55 gwei	0.0387 gwei	~1,420x

Ethereum and Base look correct only because they sit near the mixture's median.

Fix

Key the history by chainId so each chain's median is computed from its own samples. No change to the buffering or outlier logic itself.

Tests

New gas-manager.service.test.ts covers both directions:

• an expensive chain polled rarely alongside a frequently polled cheap chain keeps its own price
• a cheap chain polled rarely alongside a frequently polled expensive chain keeps its own price
• repeated samples for a single chain still smooth as before

Verified the tests fail on the previous behaviour, reproducing the exact production value:

AssertionError: expected 2751100000000n to be 6600000n

Follow-up worth considering (not in this PR)

The currentToMedianRatio < 50n branch treats a genuinely low price as "stale/broken". Now that history is per chain that is much safer, but a real sustained drop (as Avalanche had post-ACP-176) would still be discarded for a while. Worth revisiting separately.