Optimize InputSource handling with Cow and add benchmarks

[mre]

Out of curiosity, I finally tackled this TODO from the code:

// TODO: consider using Cow (to avoid one .clone() for String types)

I took a look and it turns out it doesn't touch as many places as I thought it would, so I went with it.

The result is lesser allocations for places where we handle raw input strings.

I did add a micro-benchmark for input content creation and it turns out to be decently fast.

On my machine:

• Extract from large docs: 28% improvement (21.4ms vs 29.3ms baseline)
• InputContent::from_string: 28% improvement (5.3µs vs 7.4µs baseline)
• InputContent::from_owned_string: 9% improvement (14.2µs vs 15.7µs baseline)
• from_string with owned baseline: 14% improvement (8.1µs vs 9.4µs baseline)

Not earth-shaking, but we're already operating at a decent level of performance by now, so I take it.

The question is: what do y'all think? Does it make the code too complex or is the complexity manageable? Feedback very welcome!

[thomas-zahner]

I think this is an improvement as it doesn't increase complexity very much 👍

[katrinafyi]

Should this todo also be changed?

lychee/lychee-lib/src/utils/request.rs

Lines 115 to 120 in d6b16bc

	/// Truncate the source in case it gets too long
	///
	/// This is only needed for string inputs.
	/// For other inputs, the source is simply a "label" (an enum variant).
	// TODO: This would not be necessary if we used `Cow` for the source.
	fn truncate_source(source: &InputSource) -> InputSource {