Fix Falcon-Mamba long prompt resource exhaustion

[mrjoema]

Reproduction

• Apple M4 Pro (48 GB), macOS 27.0, Python 3.13.5
• mlx==0.32.0, mlx-lm==0.31.3
• mlx-community/Falcon3-Mamba-7B-Instruct-8bits
• A 1,298-token effective prompt fails with [metal::malloc] Resource limit (499000) exceeded.

Root cause and fix

Falcon-Mamba normalized delta, B, and C twice, and each call allocated an mx.ones scale. The lazy graph therefore accumulated six temporary scales per token per layer and exceeded Metal's resource-count limit.

Normalize each value once with weight=None. The new code also replaces the nested map/lambda expression with an explicit split and three normalization calls for readability.

Verification

• Full model tests: 78 passed, 1 skipped
• Real-model prompts of 1,909 and 4,105 effective tokens complete successfully
• A regular 50-token greedy generation exactly matches the pre-fix output

Fixes #1637

[michalk8]

I tested this locally with a prompt The quick brown fox jumps over the lazy dog. " * 2000, this PR fixes it.
LGTM, thanks a lot!