fix(utils): Handle toJSON methods that return circular references

[lforst]

This PR handles a case in normalize() when a .toJSON method on an input value either returns an object with a circular reference or the input value itself, causing problems down the line.

Changes

• Normalize the return value from toJSON
• Add tests to check for edge cases

Compromise

To detect circular references we need to do a memoize call before normalizing the return value from toJSON. However, at the location we're currently calling the toJSON method, value could still be a primitive like string or number and we can't call memoize with primitives because the underlying WeakSet only takes objects.

As a compromise, we're moving the toJSON call down to where value is certainly an object or array and after the circ-ref-check. I believe it is ok for us to only call toJSON on values that are non-primitives since you can't even set a field on strings, numbers, symbols, etc. unless you modify their prototype.

If somebody has a better idea on how to tackle this best, feel free to make suggestions! - I myself am not too happy with this solution.

[github-actions]

size-limit report 📦

Path	Size
@sentry/browser - ES5 CDN Bundle (gzipped + minified)	19.3 KB (+0.03% 🔺)
@sentry/browser - ES5 CDN Bundle (minified)	59.74 KB (+0.03% 🔺)
@sentry/browser - ES6 CDN Bundle (gzipped + minified)	17.91 KB (+0.03% 🔺)
@sentry/browser - ES6 CDN Bundle (minified)	52.68 KB (+0.03% 🔺)
@sentry/browser - Webpack (gzipped + minified)	19.68 KB (+0.06% 🔺)
@sentry/browser - Webpack (minified)	64.06 KB (+0.04% 🔺)
@sentry/react - Webpack (gzipped + minified)	19.7 KB (+0.05% 🔺)
@sentry/nextjs Client - Webpack (gzipped + minified)	43.86 KB (+0.03% 🔺)
@sentry/browser + @sentry/tracing - ES5 CDN Bundle (gzipped + minified)	25.63 KB (+0.02% 🔺)
@sentry/browser + @sentry/tracing - ES6 CDN Bundle (gzipped + minified)	23.91 KB (+0.03% 🔺)

[Lms24]

This LGTM. Left just two nits.
W.r.t only calling toJSON on objects/arrays and not on primitives: If we have tests that test normalization of primitives (which I think we do), I think this is fine.
(And if it turns out to cause problems, we can still revisit this)