apparent_encoding should be cached since chardet can be slow

We have some scraper code that sometimes gets back PDFs and other times gets back HTML. Today we learned that if you access r.text in a large-ish PDF (40MB), chardet is called, which uses a lot of CPU (and a ton of memory):

r = requests.get(some_url)
r.text

That's more or less fine (best not to try to get the text of a PDF this way), but if you access r.text more than once, chardet gets run over and over.

We have code like this that performs horribly:

r = requests.get(some_url)
if bad_text in r.text:
    continue
if other_bad_text in r.text:
    continue
# ...many more tests...

When you access r.text, it checks if the encoding can come from the HTTP headers. If not, it runs the apparent_encoding property, which looks like:

    @property
    def apparent_encoding(self):
        """The apparent encoding, provided by the charset_normalizer or chardet libraries."""
        return chardet.detect(self.content)["encoding"]

I think that property should probably be cached since it's slow, so that repeated calls to r.text don't hurt so badly.

Expected Result

I expected the calls to the text property to only calculate the encoding once per request.

Actual Result

Each call to the text property re-calculates the encoding, which is slow and uses a lot of memory (this is probably a bug in chardet, but it uses hundreds of MB on a 40MB PDF right now).

System Information

$ python -m requests.help

{
  "chardet": {
    "version": "5.0.0"
  },
  "charset_normalizer": {
    "version": "2.0.12"
  },
  "cryptography": {
    "version": "36.0.2"
  },
  "idna": {
    "version": "2.10"
  },
  "implementation": {
    "name": "CPython",
    "version": "3.10.7"
  },
  "platform": {
    "release": "5.4.209-116.363.amzn2.x86_64",
    "system": "Linux"
  },
  "pyOpenSSL": {
    "openssl_version": "101010ef",
    "version": "20.0.1"
  },
  "requests": {
    "version": "2.28.1"
  },
  "system_ssl": {
    "version": "101010ef"
  },
  "urllib3": {
    "version": "1.26.5"
  },
  "using_charset_normalizer": false,
  "using_pyopenssl": true
}

Start at the Response.apparent_encoding property and its use from the r.text property. Reproduce repeated text accesses with content that requires detection, then inspect the existing response tests for a suitable regression test. Done means repeated accesses calculate the apparent encoding only once per response.